0xBakeer/TandemLLM-Qwen3.8-27B-EXL3-2.5
- Reference
- BF16 Qwen/Qwen3.8-27B
- Eval data
- wikitext-2
- Context
- not stated
- Tool
- exllamav3 1.6.0 (quantization); speed on DGX Spark in TandemLLM 0.6.2
Card snapshot main · updated 2026-10-13
Read this first
- These are claims, not our measurements. Each publisher ran its own test. A lower KLD on one card is not evidence of a better file than a higher KLD on another.
- Methods rarely match. Publishers use different reference models (BF16, their own Q8_0, someone else's Q8_0), different text, context lengths and tools. Each row shows its reference; the cards below give the rest.
- Sizes mean different things. Sizes are as published on each card: download size, GPU memory or a single file. GiB figures are converted to GB. Vision and MTP weights are included by some publishers and not by others; each row says which size it is.
Compare quants inside one measurement
Every point in a set was measured the same way by the same party, so points in one set can be compared with each other; points from different sets cannot. Your ticks carry across sets, so you can follow a quant through every set that measured it. Quants that publish no comparable numbers are listed under Repos without extracted claims.
When one publisher measured several
These are the only tables where different publishers' files went through the same test. They are still one party's measurement, often of competitors.
The ledger
| Repository | Variant | Format | Size GB | Size is | Mean KLD | Top-1 | PPL | PPL change vs the card's baseline | Other metrics (as published) | Reference | Source line |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 0xWhiteMage/Kearuga Quality-200 score is out of 180 [L65]mean_kld is Fidelity-40 mean KL, held-out is full-vocab KL [L59,L63]size is weight footprint across 3 shards + MTP head [L58] | Qwen3.8 27B Kearuga (hybrid quantized) | other | 24.85 | download | 0.0165 | 40 / 40 (100%) | — | — | kld:held-out 0.0208; top1:held-out 95.0% | vs BF16 | L58 @ main |
| 1bit-MONSTER/Q4_0-H32-GGUF [L42] | this file, W4A4 (Q4_0-H32 Hadamard-rotated) | GGUF | — | 0.055 | 89.3% | 6.096 | — | — | vs BF16 | L42 @ main | |
| 1bit-MONSTER/UD-Q4_K_XL-MTP Unsloth's Dynamic 2.0 Q4_K_XL bundled in this repo, measured by card author [L26, L34]Reference for KLD/top-1 only called 'the reference' [L34]IQ4_XS stated as 2.3x the KLD, no absolute value [L35] | Qwen3.8-27B-UD-Q4_K_XL.gguf | other | — | 0.0081 | 95.3 | — | — | — | other or no reference | L34 @ main | |
| 6block/GGUF | Q8_0 | GGUF | 28.05 | download | — | — | 6.7900 ± 0.07391 | — | — | vs BF16 | L60 @ main |
| 6block/GGUF | Q6_K | GGUF | 22.56 | download | — | — | 6.7951 ± 0.07403 | — | — | vs BF16 | L61 @ main |
| 6block/GGUF | Q5_K_M | GGUF | 20.08 | download | — | — | 6.8010 ± 0.07406 | — | — | vs BF16 | L62 @ main |
| 6block/GGUF | Q5_K_S | GGUF | 19.55 | download | — | — | 6.7973 ± 0.07396 | — | — | vs BF16 | L63 @ main |
| 6block/GGUF | Q4_1 | GGUF | 18.14 | download | — | — | 6.8105 ± 0.07411 | — | — | vs BF16 | L64 @ main |
| 6block/GGUF | Q4_K_M | GGUF | 17.75 | download | — | — | 6.8103 ± 0.07414 | — | — | vs BF16 | L65 @ main |
| 6block/GGUF | IQ4_NL | GGUF | 17.04 | download | — | — | 6.8491 ± 0.07484 | — | — | vs BF16 | L66 @ main |
| 6block/GGUF | Q4_K_S | GGUF | 16.84 | download | — | — | 6.8291 ± 0.07440 | — | — | vs BF16 | L67 @ main |
| 6block/GGUF | Q4_0 | GGUF | 16.77 | download | — | — | 6.8749 ± 0.07520 | — | — | vs BF16 | L68 @ main |
| 6block/GGUF | IQ4_XS | GGUF | 16.41 | download | — | — | 6.8529 ± 0.07487 | — | — | vs BF16 | L69 @ main |
| 6block/GGUF | Q3_K_M | GGUF | 14.21 | download | — | — | 7.1020 ± 0.07902 | — | — | vs BF16 | L70 @ main |
| 6block/GGUF | IQ3_M | GGUF | 13.49 | download | — | — | 7.0306 ± 0.07531 | — | — | vs BF16 | L71 @ main |
| 6block/GGUF | Q3_K_S | GGUF | 13.06 | download | — | — | 7.2176 ± 0.08074 | — | — | vs BF16 | L72 @ main |
| 6block/GGUF | IQ3_XXS | GGUF | 12.46 | download | — | — | 7.1994 ± 0.07848 | — | — | vs BF16 | L73 @ main |
| 6block/GGUF | Q2_K | GGUF | 11.93 | download | — | — | 7.4700 ± 0.08388 | — | — | vs BF16 | L74 @ main |
| 6block/GGUF | IQ2_M | GGUF | 11.27 | download | — | — | 7.5101 ± 0.08176 | — | — | vs BF16 | L75 @ main |
| 6block/GGUF | IQ2_XS | GGUF | 10.59 | download | — | — | 7.9668 ± 0.08803 | — | — | vs BF16 | L76 @ main |
| 6block/GGUF PPL cells carry +/- error bars [L57-77] | IQ2_XXS | GGUF | 10 | download | — | — | 8.5638 ± 0.09560 | — | — | vs BF16 | L77 @ main |
| abhishekchohan/AWQ-INT4-FP8KV weights approx 20 GB VRAM; total approx 33.5 GB c38 at 262k ctx [L104-107]recovery % vs AWQ-INT4: c4,c8,c12,c16,c20,c24,c28,c32 [L69-76]all deltas within sampling noise [L78]needle-in-a-haystack 9/9 [L88] | Qwen3.8-27B-AWQ-INT4-FP8KV (this model) | AWQ | 20.00 | GPU weights | — | — | — | — | GSM8K 95.91; GPQA Diamond 60.61; Minerva Math500 79.00; ARC-Easy 82.45; ARC-Challenge 56.66; HellaSwag 56.05; Winogrande 76.56; TruthfulQA MC1 35.74 | other or no reference | L69 @ main |
| abihsoro/AWQ-INT4 Text-only, no vision tower [L16]Qwen3.6-27B-AWQ-INT4 comparison cells (c18, c20) omitted: measured on a different base model [L41-46]GDN layers got plain group quantization, not full AWQ smoothing [L19] | Qwen3.8-27B-AWQ-INT4 | AWQ | 17.60 | download | — | — | — | — | GSM8K flexible-extract 96%; GSM8K strict-match 89% | other or no reference | L27 @ main |
| abuarchive/2bit-oQ2 Perplexity measured with +/- error [L31]Prompt battery completed 8 of 16 (7 repetition loops, 1 token-limit failure) [L35-37] | oQ2 (this model) | other | 11.00 | download | — | — | — | — | ppl:tulu-3-sft-mixture 8.124 ± 0.216; PromptBattery-completed-of-16 completed 8 | other or no reference | L29 @ main |
| airagrp/mlx-mxfp8-L KLD is mean per-token D_KL(p_bf16 || p_model) in nats [L88]ppl_delta is absolute PPL difference, not percent [L76] | Qwen3.8-27B-mlx-mxfp8-L | MLX | 37.36 | download | 0.0046 | — | 6.937 | — | — | vs BF16 | L85 @ main |
| airagrp/mlx-mxfp8-M delta PPL -0.042 (c25) is absolute, not percent; not mappedprose says ~32 GB / 9.2 bits per weight [L34] | Qwen3.8-27B-mlx-mxfp8-M | MLX | 31.98 | download | 0.0068 | — | 6.893 | — | — | vs BF16 | L84 @ main |
| airagrp/mlx-mxfp8-S KLD over full vocabulary [L89]ppl_delta is absolute, not percent [L77]prose says effective size ~29 GB (6.5 bits per weight) vs table 26.62 GiB [L35,L83] | Qwen3.8-27B-mlx-mxfp8-S | MLX | 28.58 | download | 0.0278 | — | — | — | ppl:wikitext-2 6.922 | vs BF16 | L83 @ main |
| airagrp/mlx-mxfp8-XL delta PPL vs bf16 is absolute (+0.002), not percent; c33 [L86]card prose says effective size ~39 GB, 11.2 bits per weight [L34] | Qwen3.8-27B-mlx-mxfp8-XL | MLX | 38.99 | download | 0.0037 | — | 6.937 | — | — | vs BF16 | L86 @ main |
| airagrp/mlx-nvfp4-L delta PPL +0.051 (c13) is absolute, not percent [L82]prose says effective ~24 GB, 7.9 bits per weight [L35]; table size 22.31 GiB | Qwen3.8-27B-mlx-nvfp4-L | MLX | 23.96 | unknown | 0.0407 | — | 6.987 | — | — | vs BF16 | L82 @ main |
| airagrp/mlx-nvfp4-M effective ~23 GB, 7.4 bits per weight [L35] | Qwen3.8-27B-mlx-nvfp4-M | MLX | 23.16 | download | — | — | — | — | ppl:wikitext2 6.983; kld:wikitext2 0.0490 | vs BF16 | L81 @ main |
| airagrp/mlx-nvfp4-S Delta PPL +0.088 given, not mapped [L79]Prose says ~21 GB / 5.9 bpw [L34], which differs from table 19.15 GiB | Qwen3.8-27B-mlx-nvfp4-S | MLX | 20.56 | download | 0.0573 | — | 7.024 | — | — | vs BF16 | L79 @ main |
| airagrp/mlx-nvfp4-XL table size 28.84 GiB vs prose ~31 GB (8.9 bits per weight) [L34]absolute dPPL (+0.089) not mapped, key needs percent | Qwen3.8-27B-mlx-nvfp4-XL | MLX | 30.97 | download | 0.0382 | — | — | — | ppl:wikitext2 7.024 | vs BF16 | L83 @ main |
| aj9o9/GGUF KLD on wiki.test.raw [L87,L94]hard 5/6 not mapped | Qwen3.8-27B-gdn8-q6attn-iq3ffn.gguf (this mix) | GGUF | 13.60 | download | 0.073 | — | — | — | GLSL 0.863 | vs BF16 | L35 @ main |
| akopytko/NVFP4-GGUF quality table row N4_0 [L66]; speed table calls this quant 'this NVFP4' [L31]; assumed same fileQ4_0 w/o imatrix 'matches the size and density of our quant byte by byte' (14.63 GiB, 4.60 BPW) [L62] | N4_0 | GGUF | — | 0.066254 +/- 0.000702 | 88.525 +/- 0.141% | — | — | — | vs BF16 | L66 @ main | |
| Alogotron/MixQ5A6E8-GGUF Frontier table L99-106; KL/PPL from Vulkan lane, throughput from CUDA lane [L95-97]Card disclaims best/recommendation claims [L115] | MixQ5A6E8-v0 | GGUF | 20.80 | download | 0.003083 | — | 6.6674 | — | — | vs BF16 | L104 @ main |
| amd/Quark-AWQ-INT4-W4A16 recovery % column (c3, c6, c9, c12) not mapped; ratio of quantized to base [L36]GSM8K rows are thinking vs non-thinking mode [L29-30] | Qwen3.8-27B-Quark-AWQ-INT4-W4A16 | AWQ | — | — | — | 8.8250 ppl:wikitext | — | BFCL_single_turn 24.06% | vs BF16 | L29 @ main | |
| amd/Quark-Qronos-INT4-W4A16 recovery % for perplexity is inverted base/quant [L36]no model size stated | Qwen3.8-27B-Quark-Qronos-INT4-W4A16 | INT4 | — | — | — | 8.6819 ppl:wikitext | — | GSM8K-thinking-flexible 94.62; GSM8K-thinking-strict 94.69; GSM8K-nothink-flexible 89.99; GSM8K-nothink-strict 89.84 | vs BF16 | L29 @ main | |
| andrew-paul/GGUF 28Q is a screening metric, % of BF16, lower primary run used [L22]files include Q3_K MTP weights [L13] | KERN-eqQ4 | GGUF | 14.73 | download | 0.030233 | — | — | — | — | vs BF16 | L17 @ main |
| andrew-paul/GGUF | KERN-eqQ3_M | GGUF | 13.04 | download | 0.046237 | — | — | — | — | vs BF16 | L18 @ main |
| andrew-paul/GGUF | KERN-eqQ3_S | GGUF | 11.59 | download | 0.080428 | — | — | — | — | vs BF16 | L19 @ main |
| andrew-paul/GGUF | KERN-eqIQ2_M | GGUF | 9.6 | download | 0.164745 | — | — | — | — | vs BF16 | L20 @ main |
| anemll/anemll-forge- KL-512 trace, BF16 ppl 2.136 [L64]; M5 Max KL-512 0.0517 (c9) [L63] | ANEMLL release 0.2 (ANE M6, compiled) | other | — | 0.0518 | 91.9% | 2.131 | — | kld_99 0.50; ppl:verify_path 6.052; ppl:8k 7.552; ppl:64k 4.537 | vs BF16 | L60 @ main | |
| anubhav200/exl3-4.00bpw-bf16vision turboderp's EXL3 4.00 bpw quant unchanged plus bf16 vision tower [L19-20]NLL on reference-generated tokens, slightly favourable to EXL3 [L75-76] | this (EXL3 4.00 bpw) | EXL3 | — | — | 93.50% | — | — | — | vs BF16 | L80 @ main | |
| Ar3ntiz/qwen3_27B_side8 top-1 given as fraction 0.95 [L49]answer-level scores are counts correct out of 20 per subset [L60-64]BBH and Long context columns not in candidates for this row [L64]bpw 4.87 stated in prose only [L19] | SUC1 side8 (this model) | other | 16.36 | download | 0.023 | 0.95 | 7.15 | — | GSM8K 17; MATH 13; MMLU-Pro 8 | vs BF16 | L49 @ main |
| armin1/NVFP4-FP8-Mixed-LH size from prose '23.4 GB on disk' [L53]DeltaNLL +0.0111 [L89] not a mapped keyGSM8K thinking on, 16K cap [L94,L99] | This model | NVFP4 | 23.40 | download | 0.0178 | 95.98% | — | — | kld_99 0.206; GSM8K 96.06% | vs BF16 | L53 @ main |
| armin1/NVFP4-FP8-Mixed-LH-Heretic Size 22.6 GB on disk from prose [L78]DeltaNLL +0.0117 not mapped (no key) [L111]Abliterated; Heretic KL 0.0638 and refusals 28/100 measure uncensoring, not quant fidelity [L71, L81] | NVFP4-FP8-Mixed-LH-Heretic (this model) | NVFP4 | 22.60 | download | 0.0188 | 95.81% | — | — | kld_99 0.217; GSM8K 95.83% | vs BF16 | L78 @ main |
| atakans/NVFP4-BF16head Size/speed measured on DGX Spark [L33-38]BF16 lm_head adds 1.7 GiB [L41] | This model | NVFP4 | 23.62 | download | 0.0295 | 91.1% | — | 1.022 | kld_99 0.309 | vs BF16 | L30 @ main |
| AtomicChat/GGUF [L55] | Q8_0 | GGUF | 28.90 | download | 0.00064 | 98.92% | — | — | — | vs BF16 | L55 @ main |
| AtomicChat/GGUF [L56] | AD-Q6_K | GGUF | 25.00 | download | 0.00107 | 98.67% | — | — | — | vs BF16 | L56 @ main |
| AtomicChat/GGUF [L57] | AD-Q6_K-Q5_K | GGUF | 23.10 | download | 0.00252 | 97.94% | — | — | — | vs BF16 | L57 @ main |
| AtomicChat/GGUF [L58] | AD-Q5_K_M | GGUF | 20.20 | download | 0.00419 | 97.34% | — | — | — | vs BF16 | L58 @ main |
| AtomicChat/GGUF [L59] | AD-Q5_K_M-Q4_K_M | GGUF | 18.60 | download | 0.00730 | 96.43% | — | — | — | vs BF16 | L59 @ main |
| AtomicChat/GGUF [L60] | AD-Q4_K_M | GGUF | 17.10 | download | 0.01126 | 95.59% | — | — | — | vs BF16 | L60 @ main |
| AtomicChat/GGUF [L61] | AD-IQ4_XS | GGUF | 16.50 | download | 0.01248 | 95.39% | — | — | — | vs BF16 | L61 @ main |
| AtomicChat/GGUF [L62] | AD-IQ4_XS-IQ3_S | GGUF | 14.40 | download | 0.02660 | 93.15% | — | — | — | vs BF16 | L62 @ main |
| AtomicChat/GGUF [L63] | AD-IQ3_S | GGUF | 13.80 | download | 0.03247 | 92.41% | — | — | — | vs BF16 | L63 @ main |
| AtomicChat/GGUF [L64] | AD-IQ3_S-IQ3_XXS | GGUF | 13.00 | download | 0.04337 | 91.33% | — | — | — | vs BF16 | L64 @ main |
| AtomicChat/GGUF [L65] | AD-IQ3_XXS | GGUF | 12.10 | download | 0.06972 | 89.13% | — | — | — | vs BF16 | L65 @ main |
| AtomicChat/GGUF [L66] | AD-IQ2_S | GGUF | 11.10 | download | 0.09832 | 87.18% | — | — | — | vs BF16 | L66 @ main |
| AtomicChat/GGUF [L67] | AD-IQ2_S-IQ2_XS | GGUF | 10.20 | download | 0.13807 | 84.77% | — | — | — | vs BF16 | L67 @ main |
| AtomicChat/GGUF [L68] | AD-IQ2_XS | GGUF | 9.9 | download | 0.16170 | 83.48% | — | — | — | vs BF16 | L68 @ main |
| AtomicChat/GGUF [L69] | AD-IQ2_XXS | GGUF | 9 | download | 0.25663 | 79.44% | — | — | — | vs BF16 | L69 @ main |
| AtomicChat/GGUF [L70] | AD-IQ1_M | GGUF | 8.5 | download | 0.34212 | 76.34% | — | — | — | vs BF16 | L70 @ main |
| AveLabs/Perplexing-Pegasus-NVFP4 Chat benchmarks vs stock NVFP4 under identical SGLang settings [L84-91]IFEval row is prompt-level strict / loose [L89] | Perplexing-Pegasus NVFP4 (chat) | NVFP4 | — | — | — | — | — | GSM8K 87.5%; MMLU-Pro 81.6% | vs BF16 | L89 @ main | |
| AveLabs/Perplexing-Pegasus-NVFP4 Decision accuracy on identical requests vs original BF16 pplx-decider [L62-69]Same-answer rate and mean prob. difference omitted (no matching metric key) [L66-69]Upstream-scoring JevBench 199-200/231 also reported [L77] | Perplexing-Pegasus NVFP4 (decisions) | NVFP4 | — | — | — | — | — | WANLI 68.7% | vs BF16 | L66 @ main | |
| Avesed/INT4-W4A16 GSM8K 96.4% (241/250) in prose; MMLU-Pro not run [L43,L51]~20 GB size in prose without a candidate id [L23] | INT4-W4A16 | INT4 | — | — | — | — | — | GSM8K 96.4 % | other or no reference | L43 @ main | |
| Avesed/INT8-W8A8 GSM8K 96.8% (242/250) in prose; MMLU-Pro not run [L35,L37]~30 GB size in prose without a candidate id [L22] | INT8-W8A8 | INT8 | — | — | — | — | — | GSM8K 96.8 % | other or no reference | L35 @ main | |
| avlp12/Alis-MLX-4bit exact full-vocab KL to bf16, pooled en+ko+code [L365-375] | 4-bit AWQ + graded (this repo) - exact KL | MLX | 15.41 | download | 0.0592 | 92.4% | — | — | — | vs BF16 | L369 @ main |
| avlp12/Alis-MLX-4bit PPL measured on pre-graded weights, not re-run on graded build (upper bound) [L349]size from at-a-glance table [L36] | 4-bit AWQ + graded (this repo) - PPL | MLX | 15.40 | download | — | — | — | — | ppl:en 5.8450; ppl:ko 6.2609; ppl:code 1.8105 | vs BF16 | L36 @ main |
| avlp12/Alis-MLX-6bit PPL table measured on pre-graded weights, not re-run [L308-310]KL 0.00499 / top-1 97.5% is the graded build, exact full-vocab KL over ~100K en+ko+code tokens [L316]graded-allocation before/after table (L274-279) has unlabeled metric; omittedstated 6.501 bits per weight in prose [L250] | 6-bit (this repo) | MLX | 21.90 | download | 0.00499 | 97.5 | — | — | ppl:en 5.7924; ppl:ko 6.1018; ppl:code 1.6854 | vs BF16 | L30 @ main |
| avlp12/Alis-MLX-6bit KL 0.00311 in prose [L319]; compared against a third-party 6-bit [L319-323] | 6-bit (this repo), WikiText-2 ctx 1024 x35 windows | MLX | — | 0.00311 | 97.54 | — | — | — | vs BF16 | L320 @ main | |
| avlp12/Alis-MLX-8bit PPL measured on pre-graded weights, not re-run [L303]KL/top-1 vs bf16 from exact-KL sweep [L318] | 8-bit (this repo) | MLX | 27.90 | download | 0.00105 | 98.45 | — | — | ppl:en 5.7760; ppl:ko 6.0987; ppl:code 1.6815 † | vs BF16 | L29 @ main |
| avyukth/AWQ-INT4 GSM8K is 39/40 on a 40-problem subset, not the full test set [L75,L79]Size on disk 18 GB and weights in VRAM 17.68 GiB in separate table [L38-39]No base comparison reported | AWQ INT4 (compressed-tensors) | AWQ | — | — | — | — | — | GSM8K 97.5 | reference not stated | L75 @ main | |
| ayushcluster00/Heretic-Abliterated-Uncensored-GGUF-Config KL vs base measures abliteration damage, not quantization [L119,L126-130]refusals 0-1/100 [L130]file sizes in L136-163 are not tied to any quality metric | RVN (this repo) | GGUF | — | 0.0085 | — | — | — | — | other or no reference | L130 @ main | |
| baa-ai/RAM-24GB-MLX MMLU is a 100-question smoke test, one question is within noise [L49]in-memory footprint ~24 GiB [L31]canary: 24 SOPs x 3 seeds = 144 paired events [L64] | RAM-24GB (mixed, 7.52 bpw) | MLX | 25.70 | download | — | — | — | — | canary_invented_x 0.000; canary_recall 0.997; canary_branch 0.986 | vs BF16 | L33 @ main |
| bernhardbrieger/GPTQ-Int4 absolute perplexity on one 260-token sample, sanity check only [L237]card states no base-vs-quant benchmark or diff [L270-273]also reports 3/3 functional smoke and 6/6 vision probes [L242-264] | Qwen3.8-27B-GPTQ-Int4 | GPTQ | 20.98 | download | — | — | 2.729 | — | — | other or no reference | L34 @ main |
| bernhardbrieger/NVFP4 absolute perplexity on single 260-token sample, n=260 tokens [L223,c11]size 20.59 GB from safetensors index (19.7 main + 0.85 MTP) [L37-38]bf16 55.56 GB reference size only, no bf16 quality number [L37,L47-50] | Qwen3.8-27B-NVFP4 | NVFP4 | 20.59 | download | — | — | 2.799 | — | — | other or no reference | L37 @ main |
| bg-digitalservices/ROCmFP4-STRIX-GGUF same-top is share of tokens where quant picks same top token as F16 [L42]base type 4.49 bpw label at L34 not tied to this file's total bpw [L34] | this file (ROCmFP4 STRIX recipe, LM head Q6_K) | GGUF | 15.38 | download | — | — | — | — | kld:EN 0.034788; kld:DE 0.033229; top1:EN 91.43% | other or no reference | L46 @ main |
| billmccartney/exl3-8.00bpw size is safetensors shard bytes [L195]; mean KL unadjusted, no BF16 self-noise pass [L193-194] | EXL3 8.0bpw | EXL3 | 29.34 | download | 0.00142020 | — | — | — | ppl:wikitext2_r100 6.94884191 | vs BF16 | L38 @ main |
| bowmanslayer/W4A16-vision eval table does not say which variant (-vision or -vision-mtp) was run; both from same quantization run [L45]thinking ON scores not comparable to leaderboards with thinking off [L109-L111]truncation rates reported separately [L119-L130] | W4A16-vision | INT4 | 18.00 | download | — | — | — | — | MMLU 84.00; MMLU_ex_truncation 95.45; CMMLU 84.67; CMMLU_ex_truncation 92.03; C-Eval 78.67; C-Eval_ex_truncation 88.06; ARC-C 94.00; ARC-C_ex_truncation 94.63; TruthfulQA-MC1 85.33; TruthfulQA-MC1_ex_truncation 92.09; GSM8K 97.00; GSM8K_ex_truncation 98.98; MATH-500 77.00; MATH-500_ex_truncation 79.38; BBH 88.67; BBH_ex_truncation 97.01; HumanEval 82.32; HumanEval_ex_truncation 91.22; IFEval-strict 79.00; IFEval-strict_ex_truncation 91.57 | other or no reference | L42 @ main |
| btbtyler09/GPTQ-4bit size includes BF16 vision encoder and MTP head; text decoder alone ~19 GB [L79]calibration: evol-codealpaca + C4, 256 samples [L65-66]Qwen3.6-27B vanilla GPTQ +8.45% / FOEM +1.95% figures are a different model, omitted [L61] | GPTQ 4-bit FOEM (this) | GPTQ | 21.00 | download | — | — | 6.4982 | +0.81% | — | vs BF16 | L77 @ main |
| btbtyler09/GPTQ-8bit Degradation cell reads -0.02% (within noise) [L82]Size includes BF16 vision encoder and MTP head [L73] | GPTQ 8-bit (this) | GPTQ | 31.00 | download | — | — | — | -0.02% (within noise, effectively lossless) | ppl:wikitext-2 6.4446 | vs BF16 | L70 @ main |
| Bucoid/Heretic-Ara-16GB-VRAM-IQ4-XS-MTP-GGUF kld_99 is the 99.9% KL quantile [L37]size 13 GiB, 13.3 GiB with MTP [L31]same table duplicated in English L62-71 | IQ4_XS-3.0 | GGUF | 13.96 | download | 0.027832 ± 0.000324 | 92.867% ± 0.067% | 7.046980 ± 0.045498 | — | kld_99 1.162850 | vs BF16 | L31 @ main |
| Bucoid/Heretic-Ara-16GB-VRAM-IQ4-XS-MTP-GGUF older release; kld_99 is 99.9% quantile [L37] | IQ4_XS-2.0 | GGUF | 13.64 | download | 0.033398 ± 0.000308 | 91.619% ± 0.072% | 7.102940 ± 0.046017 | — | kld_99 1.130034 | vs BF16 | L31 @ main |
| Bucoid/IQ4-XS-MTP-16GB-VRAM-GGUF kld_99 is the 99.9% KL quantile [L25]table says GB but prose says GiB for 12.9 [L9] | IQ4_XS | GGUF | 12.90 | download | 0.03168 ± 0.00029 | 91.845% ± 0.071% | 7.0875 ± 0.0459 | — | — | vs BF16 | L19 @ main |
| Bucoid/Uncensored-IQ4-XS-MTP-16GB-VRAM-GGUF table cell says GB, prose says 12.9 GiB [L22,L52,L60]English table omits own max KLD, mean dp, RMS dp values; taken from Chinese table [L35,L38,L39]99.9% KL quantile, not 99.0% | IQ4_XS | GGUF | 12.90 | download | 0.03268 ± 0.00030 | 91.655% ± 0.072% | 7.1481 ± 0.0465 | — | kld_max 16.017（更小）; kld_999 1.075 | other or no reference | L35 @ main |
| cafonez/ROCmI4-MTP-GGUF perplexity corpus not stated [L206]W4A4 path adds ~5.4% PPL on a 25-chunk sample [L208-209] | ROCmI4 artifact | GGUF | — | — | — | 6.3068 | +6.1% | HumanEval 94.5; HumanEval+ 91.5 | other or no reference | L198 @ main | |
| caiovicentino1/GPTQ-gs64-FOEM tiny sub-samples, indicative only [L97]no BF16 baseline [L80]4-bit gs64 with fp16/bf16 skip-list layers [L37] | Qwen3.8-27B-GPTQ-gs64-FOEM | GPTQ | 27.29 | download | — | — | — | — | ppl:C4 9.90; HellaSwag 0.35; ARC-Challenge 0.45; MMLU 0.47; GSM8K 0.75 | other or no reference | L37 @ main |
| Capicua25x/MXFP4-Quark-RDNA4 Accuracy table L171-184; +/-2 items noise band [L169]Throughput tables are speed only and omitted [L106-157] | C (this model) MXFP4 @ bf16 KV | MXFP4 | — | — | — | — | — | GSM8K think flex 0.98; GSM8K think strict 0.96; IFEval inst 0.95; IFEval prompt 0.91; GPQA-Diamond 0.92; tau2-telecom 0.904; tau2-airline 0.840; tau2-retail 0.767 | vs BF16 | L173 @ main | |
| cdiamond/iMatrix-NVFP4-MTP-GGUF short WikiText-2 control; author says gap to Q4_1 is within error [L55] | iMatrix NVFP4 MTP GGUF (mixed precision) | GGUF | 17.10 | download | — | — | 6.1197 | — | — | other or no reference | L24 @ main |
| Chungulus/EXL3 top-1 agreement reported as fraction 0.943359, not a percent [L50]top-5 overlap equals the same 0.943359 [L51]; median KL 0.00308530, P95 KL 0.02567411 and logit cosine 0.99754137 also reported, not mappedonly 512 token positions evaluated [L42]quantizer config states bits 4.0 and head_bits 6.0 [L21] | Qwen3.8-27B EXL3 (bits 4.0) | EXL3 | 19.11 | download | 0.00759337 | 0.943359 | — | -1.2346% | — | reference not stated | L22 @ main |
| Chungulus/FP8 Median KL c2, P95 KL c4, top-5 set overlap c7, logit cosine c8 [L47-52]; top-1 value printed as fraction 0.968750 [L50]Only 64 token positions evaluated [L42]Artifact size 29.127 GiB (c9) in provenance prose, not in metrics table [L22] | Qwen3.8-27B FP8 | FP8 | — | 0.00575227 | 0.968750 | — | -0.5484% | — | reference not stated | L46 @ main | |
| Chungulus/GPTQ-INT4-G128 Top-1 agreement printed as fraction 0.935547 not percent [L50]Median KL 0.00990394 and P95 KL 0.08463629 not mapped to a metric key [L47-48]Top-5 overlap 0.912109 and logit cosine 0.99503607 not mapped [L51-52]Reference model for KL/PPL not named in digest [L40-44] | GPTQ-INT4-G128 | GPTQ | 19.58 | download | 0.02518878 | 0.935547 | — | +4.0345% | — | reference not stated | L22 @ main |
| Chungulus/GPTQ-INT4-G32 top1 and top5 overlap given as fractions (0.93...), not percent [L50-L51]size is artifact size from prose, not the metrics table [L22]reference model not named in validation section [L40-L44] | GPTQ-INT4-G32 | GPTQ | 21.01 | download | 0.01732363 | 0.933594 | — | +3.2276% | — | reference not stated | L22 @ main |
| Chungulus/GPTQ-INT8 Top-1 agreement printed as fraction 0.978516 not percent [L50]P95 KL 0.00221081 not mapped (key is kld_99) [L48]Artifact size 29.655 GiB [L22]; reference model not named, 512 positions only [L42] | Qwen3.8-27B GPTQ-INT8 | GPTQ | 31.84 | download | 0.00079757 | — | — | +0.1411% | — | reference not stated | L22 @ main |
| Chungulus/MLX-3bit-Group32 top1_token_agreement is a fraction (0.83), not percent [L128]BF16 reference perplexity 9.847 [L125]; only 106 positions [L122]mean KL 0.0918 given in prose only [L124]artifact size 15.266 GB decimal [L38] | Qwen3.8-27B MLX 3-bit group32 | MLX | 15.27 | download | 0.09177716048258655 | — | perplexity": 10.580787393174923, | — | — | vs BF16 | L38 @ main |
| Chungulus/MLX-4bit quality proxy is mean semantic similarity, stated as not ground-truth accuracy [L55]exact_matches 3 [L106]acceptance_rate and tps are speed-side, not mapped [L94-95] | Qwen3.8-27B-MLX-4bit | MLX | 16.95 | download | — | — | — | — | mean_semantic_similarity 0.948318076133728 | reference not stated | L38 @ main |
| Chungulus/MLX-4bit-Group32 top1 value is a fraction (0.943), not percent [L128]reference BF16 perplexity 9.847 [L125]; perplexity delta 0.198 [L127]very small sample: 106 positions [L122] | MLX affine 4-bit group size 32 | MLX | 18.63 | download | 0.01875024640335227 | 0.943396226415094 | perplexity": 10.045437946476026, | — | — | vs BF16 | L38 @ main |
| Chungulus/MLX-5bit-Group128 BF16 reference ppl 9.8473 (c8), candidate ppl lower than reference, delta -0.0830 (c10) [L125-127]top1 agreement is a fraction 0.9906, not percent [L128]Only 106 positions scored [L122] | MLX-5bit-Group128 | MLX | 19.47 | download | 0.007879387376443395 | 0.990566037735849 | perplexity": 9.764229178940486, | — | — | vs BF16 | L38 @ main |
| Chungulus/MLX-5bit-Group32 top1_token_agreement is a fraction (0.9528...) not a percent [L128]perplexity_delta 0.0367 absolute, not percent [L127]Size is decimal GB artifact size [L38] | Qwen3.8-27B MLX-5bit-Group32 | MLX | 21.99 | download | — | 0.952830188679245 | perplexity": 9.883995913928073, | — | — | vs BF16 | L38 @ main |
| Chungulus/MLX-6bit-Group32 top1 value is a 0-1 fraction not a percent [L128]ppl_delta_abs is absolute perplexity difference, not percent [L127]KL measured with float16-stored BF16 log-probs, finite-storage error [L132]Very small sample (106 positions) [L122] | MLX-6bit-Group32 | MLX | 25.35 | download | 0.001391285889834728 | 0.990566037735849 | perplexity": 9.812861698114856, | — | mean_semantic_similarity 0.9533392310142517 | vs BF16 | L38 @ main |
| Chungulus/MLX-8bit-Group32 top1 is a fraction 0.9811, not percent [L128]only 106 positions [L122]perplexity_delta 0.0060 absolute (c10) [L127] | Qwen3.8-27B MLX-8bit-Group32 | MLX | 32.07 | download | 0.0003183703201840509 | 0.981132075471698 | perplexity": 9.853291632468812, | — | — | vs BF16 | L38 @ main |
| Chungulus/MLX-MXFP8 semantic similarity described as a measured proxy, not ground-truth accuracy; reference not stated [L55,L105]exact_matches 4 [L106]size 29.553 GB decimal [L38] | Qwen3.8-27B MLX-MXFP8 | MLX | 29.55 | download | — | — | — | — | mean_semantic_similarity 0.9667035639286041 | reference not stated | L38 @ main |
| Chungulus/MLX-oQ4e-MTP Only a proxy is reported: mean semantic similarity 0.899, 4 exact matches, described as a measured proxy not ground-truth accuracy [L55, L144-145]Artifact 16.995 GB decimal [L38]; peak memory 17.79 GB, average generation 21.8 tok/s [L147-148] | Qwen3.8-27B MLX-oQ4e-MTP | MLX | 17.00 | download | — | — | — | — | mean_semantic_similarity 0.8991085827350617 | reference not stated | L38 @ main |
| Chungulus/MLX-oQ5 top1_token_agreement printed as fraction 0.981..., not percent [L167]perplexity_delta -0.0966 (c10) is absolute not percent, omitted [L166]bpw stated only as 'mixed around 5.5 bpw' [L28]; size is decimal GB [L38] | MLX-oQ5 | MLX | 20.31 | download | 0.00603105514398161 | 0.981132075471698 | perplexity": 9.75069257534557, | — | — | vs BF16 | L38 @ main |
| Chungulus/MLX-oQ8 decimal GB artifact size [L38]only 106 positions, tiny sample [L161]bit width described as mixed around 8.6 bpw, no numeric cell [L28]mean KL only in JSON prose [L163] | Qwen3.8-27B MLX-oQ8 | MLX | 30.02 | download | 0.0007939068249092912 | — | — | — | top1_agreement_fraction 1.0 | vs BF16 | L38 @ main |
| cloudnathan5/NVFP4a16-GPTQ Checkpoint size 28.6 GB vs 55.6 GB BF16 stated in prose [L30] | Qwen3.8-27B-NVFP4a16-GPTQ (this checkpoint) | NVFP4 | — | — | — | — | +2.37% | ppl:wikitext-2 6.7129 | vs BF16 | L98 @ main | |
| cloudnathan5/NVFP4a4-FP8-GPTQ Checkpoint size stated in prose, 2.42x smaller than BF16 [L29]Latency table is speed only [L58-65] | this checkpoint | NVFP4 | 23.00 | download | — | — | — | +4.64% | ppl:wikitext2 6.8614 | vs BF16 | L29 @ main |
| cloudnathan5/NVFP4a4-GPTQ size from prose, not the perplexity table [L28] | this checkpoint (NVFP4a4 GPTQ) | NVFP4 | 28.60 | download | — | — | 6.7733 | +3.29% | — | vs BF16 | L28 @ main |
| CodeMasterCody3D/taardis-27b-full-ternary V1 listed as 7.16 GB in file table [L35] and 7.17 GB in comparison table [L63] | TAARDIS-27B-Full-Ternary-V1.gguf (weights alone) | other | 7.2 | download | — | — | 13.61 ppl:wikitext | — | — | reference not stated | L35 @ main |
| CodeMasterCody3D/taardis-27b-full-ternary V2 is a lossless repack of V1, same weights [L53-56]KV-cache ternary types cost +11.2% to +11.4% PPL, q4_0 +0.16% vs f16 KV [L235-239]; not weight-quant quality so omitted | TAARDIS-27B-Full-Ternary-V2-1.75bit.gguf (weights alone) | other | 5.9 | download | — | — | 13.6114 ppl:wikitext | — | — | reference not stated | L32 @ main |
| cometkim/nvfp4full-NInfer 3-round means with +/- spread [L226-228]Artifact is 18.07 GiB on disk; device weights 16.02 GiB no vision/no spec [L191,L199]Quality tables describe v2-era weights byte-preserved into v3 [L117] | nvfp4full .ninfer (INT8 group-64 KV, native 262,144) | NVFP4 | — | — | — | — | — | GPQA-Diamond 87.88 ± 2.62; AIME 2026 93.33 ± 3.34; LongBench v2 short 67.41 ± 1.95 | other or no reference | L226 @ main | |
| curvedinf/GPTQ-INT8-W8A8-GS128 KLD 0.011 is vs the gs32 reference, not BF16; gate 0.02 [L48]Activation quantization adds only +0.0032 KLD on top (no candidate id) [L49] | GPTQ INT8 W8A8 GS128 | GPTQ | — | 0.011 | — | — | — | — | other or no reference | L48 @ main | |
| CySpiegel/Int4-AutoRound size from prose '19 GB on disk' [L21]low GSM8K absolutes attributed to harness artifact for both sides [L49]calibration pile-10k, 4-bit group 128 [L26,L27] | This model (INT4) | AutoRound | 19.00 | download | — | — | — | — | GSM8K 5-shot 43.2 | other or no reference | L21 @ main |
| Davidmg0815/MTP-GGUF HumanEval+ 145/164 [L430]; variant MATH-500 table L466 is a different token budget than the 98.0% headline [L48]German decimal commas in source | qwen3.8-27b-Q8_0.gguf | GGUF | 273.0 | download | — | — | — | — | HumanEval+ 88,4 % | other or no reference | L48 @ main |
| Davidmg0815/MTP-GGUF HumanEval+ 148/164 [L428] | qwen3.8-27b-Q6_K.gguf | GGUF | 210.0 | download | — | — | — | — | HumanEval+ 90,2 % | other or no reference | L68 @ main |
| dbirks/NVFP4-AutoRound values shown with +- confidence intervals [L89-91]MMLU-Pro subsample not comparable with full-split numbers elsewhere [L94-95]size from prose, 20.6 GB on disk [L25] | NVFP4 W4A4 | NVFP4 | 20.60 | download | — | — | — | — | GSM8K 0.905 ±0.016; HumanEval 0.927 ±0.040; MMLU-Pro 0.796 ±0.021 | vs BF16 | L25 @ main |
| dbirks/W4A16-AutoRound Size 19.5 GB on disk from prose [L26]; effective bits/weight about 4.1 [L105]MMLU-Pro is 100 questions per subject, internal reference only [L88-89] | W4A16 AutoRound (int4) | AutoRound | 19.50 | download | — | — | — | — | GSM8K 0.917 ±0.015; HumanEval 0.957 ±0.031; MMLU-Pro 0.826 ±0.019 | vs BF16 | L26 @ main |
| ddalcu/MLX-Serve-6bit 6.50 bits per weight stated in prose [L21]Decode 46.3 tok/s with MTP on, M4 Max [L46] | 6-bit (this one) | MLX | 23.20 | download | 0.102 | 95.0% | — | — | — | vs BF16 | L46 @ main |
| ddalcu/MLX-Serve-iQ-MLX-3.8bpw resident weights as shipped [L48]12.89 GB on disk total quantized [L39] | iQ-MLX (this one) | MLX | 13.40 | resident | 0.466 | 81.6% | — | — | — | vs BF16 | L48 @ main |
| dealignai/Bonsai-2-27B-1bit-CRACK-GGUF MMLU delta vs base -1.23 pp [L107]refusal-rate tables are not quality-vs-reference metrics; omittedbpw label 2.13 [L54] conflicts with header 1.75 bpw [L33]; unclear | CRACK PTQ1_0 | GGUF | — | — | — | — | — | MMLU 38.46% | other or no reference | L54 @ main | |
| dealignai/Bonsai-2-27B-Ternary-CRACK-GGUF MMLU delta vs base -0.62 pp [L107]Refusal-rate tables (HarmBench-320, per-category, per-reasoning-mode) are not quality metric keys [L64-100]Card states 7.2 GB file size in prose [L33] | CRACK PQ2_0 | GGUF | — | — | — | — | — | MMLU 39.91% | other or no reference | L54 @ main | |
| dealignai/JANG_2D-CRACK MMLU delta -3.16pp [L118]base MMLU differs by profile: 80.0 here vs 88.77 for 4D/6D [L118-119] | JANG_2D CRACK | other | 11.00 | download | — | — | — | — | MMLU_base 80.0% | other or no reference | L118 @ main |
| dealignai/JANG_4D-CRACK CRACK abliterated weights, not a plain quant; MMLU delta -1.05pp vs base includes abliteration effect [L72,L87]base MMLU 88.77% [L119] | 4D (JANG_4D CRACK, ~4-bit affine imatrix MLX) | other | 17.00 | download | — | — | — | — | MMLU 87.72% | other or no reference | L119 @ main |
| dealignai/JANG_6D-CRACK CRACK is an abliteration; MMLU compares abliterated vs base, not quant vs bf16 [L84]HarmBench compliance is a refusal metric, not mapped [L85] | JANG_6D CRACK | other | 24.00 | download | — | — | — | — | MMLU 89.12%; MMLU_base 88.77% | other or no reference | L120 @ main |
| dealignai/MXFP8-CRACK abliteration compliance score, not a quantization fidelity metric [L77-78]L83 HarmBench row label mixes compliance and refuses; value 100.0% taken from per-quant table [L83,L119] | MXFP8 CRACK | MXFP8 | 27.00 | download | — | — | — | — | MMLU 86.67%; HarmBench-240 100.0% | other or no reference | L82 @ main |
| Derpyhue/abliterated-2.55bpw-SC-SH median KLD c9, p90 KLD c10 [L46]Perplexity (increase) column c11 left out: ambiguous metric, BF16 reference 1.3418 is not a normal perplexity [L45-46]predicted KL 0.0706 (sc_optimize estimate) c1 is not a measurement [L28]weights count quantized tensors only; about 11.81 GB total on disk [L41]measured against abliterated BF16, not comparable with main series [L19] | 2.55 H4 -sh | other | 8.4 | text weights (excl. vision, MTP) | 0.0442 | — | — | — | — | vs BF16 | L27 @ main |
| Derpyhue/abliterated-2.60bpw-SC-SH Weights (GiB) counts quantized tensors only; ~11.96 GB on disk [L41]PPL cell reads 1.4048 (+4.7%) vs BF16 reference 1.3418 [L47]; median KLD 0.0023 and p90 KLD 0.122 also reported, not mappedmeasured against the abliterated BF16 reference, not comparable with main-series tables [L37] | 2.60 H4 -sh (this repo) | other | 8.6 | text weights (excl. vision, MTP) | 0.0417 | — | 1.4048 (+4.7%) | — | — | vs BF16 | L27 @ main |
| Derpyhue/abliterated-2.65bpw-SC-SH weights GiB counts quantized tensors only; ~12.11 GB total on disk [L41]reference is the abliterated BF16, not comparable with main-series tables [L37]'Perplexity (increase)' column (1.3986 (+4.2%)) is ambiguous; not mapped [L48]sc_optimize predicted KL 0.0616 is an estimate [L28] | 2.65 H4 -sh | other | 8.8 | text weights (excl. vision, MTP) | 0.0377 | — | — | — | — | vs BF16 | L27 @ main |
| Derpyhue/exl3-2.65bpw-SC size is full repo footprint [L37]Perplexity values near 1.37 look like a ratio or relative figure; basis not stated [L39]predicted KL 0.0587 is an sc_optimize estimate, not measured [L25] | 2.65 H4 | EXL3 | 12.11 | download | 0.032903 | — | 1.3795 | — | — | vs BF16 | L24 @ main |
| Derpyhue/exl3-2.65bpw-SC-SH perplexity values near 1.37 look like a ratio or differently-based metric; basis unclear [L40-48]median KLD c12 and p90 KLD c13 not mappedsc_optimize predicted KL 0.0587 is an estimate, not a measurement [L25] | 2.65 H4 -sh | EXL3 | 12.11 | download | 0.032541 | — | 1.3739 | — | — | vs BF16 | L24 @ main |
| Derpyhue/exl3-2.67bpw-SC-SH median KLD c17, p90 KLD c18 [L44]size is full repo footprint [L38]predicted KL 0.0571 is an sc_optimize estimate, not measured [L25] | 2.67 H4 -sh | EXL3 | 12.17 | download | 0.031061 | — | 1.3748 | — | — | vs BF16 | L24 @ main |
| Derpyhue/exl3-2.70bpw-SC size is full repo footprint [L37]predicted KLD 0.0549 is an sc_optimize estimate, not measured [L25] | 2.70 H4 | EXL3 | 12.26 | download | 0.030083 | — | 1.3746 | — | — | vs BF16 | L24 @ main |
| Derpyhue/exl3-2.70bpw-SC-SH size is full repo footprint [L38]perplexity values near 1.37 as printed; basis unclear [L46] | 2.70 H4 -sh | EXL3 | 12.26 | download | 0.030040 | — | 1.3748 | — | — | vs BF16 | L24 @ main |
| Derpyhue/exl3-2.75bpw-SC size on disk is the full repo footprint [L37]median/p90 KLD not mapped (c32, c33) [L46]predicted KL 0.0514 c1 is an sc_optimize estimate, not measured [L25] | 2.75 H4 | EXL3 | 12.41 | download | 0.027894 | — | 1.3725 | — | — | vs BF16 | L24 @ main |
| Derpyhue/exl3-2.75bpw-SC-SH Bold row = this repo [L38,L48]Predicted KL 0.0514 (c1) is an sc_optimize estimate, not the measured value [L25] | 2.75 H4 -sh | EXL3 | 12.41 | download | 0.026945 | — | 1.3685 | — | — | vs BF16 | L24 @ main |
| Derpyhue/exl3-2.80bpw-SC size is full repo footprint [L37]perplexity ~1.37 is not comparable to usual PPL values; reported as printedpredicted KL 0.0481 (sc_optimize estimate) c1 omitted [L25] | 2.80 H4 | EXL3 | 12.57 | download | 0.025571 | — | 1.3690 | — | kld_median 0.001298; kld_p90 0.074292 | vs BF16 | L24 @ main |
| Derpyhue/exl3-2.80bpw-SC-SH size is full repo footprint [L38]predicted KL 0.0481 (sc_optimize estimate, c1) is not a measurement [L25] | 2.80 H4 -sh | EXL3 | 12.57 | download | 0.025098 | — | 1.3688 | — | — | vs BF16 | L24 @ main |
| deucebucket/Cerebellum-GGUF self-reported, not HF-verified [L108]size is weight-file arithmetic [L96]ppl 145 chunks ctx 512 [L117] | Qwen3.8-27B Cerebellum Q2_K Mixed 12GB | GGUF | 12.00 | download | — | — | 7.4842 +/- 0.09908 | — | — | other or no reference | L83 @ main |
| dfed24/GPTQ-GGUF reference is Q8_0, not bf16 [L38] | this file (llama.cpp vs unsloth Q8_0 reference) | GGUF | — | — | — | — | — | kld:wiki 0.0112; top1:wiki 95.70; kld:code 0.0125; top1:code 99.02; kld:multilingual 0.0084; top1:multilingual 96.40 | vs BF16 | L44 @ main | |
| drawthingsai/GGUF Size includes MTP weights; size without MTP 9.29 GiB [L78]bpw 2.96 appears only in the checkpoint name [L56] | DT-bpw2.96-IQ3_XXS | GGUF | 10.43 | download | — | — | — | — | AIME 2026 100.00%; GPQA Diamond 86.36%; IFEval 93.90%; BFCL v4 (1k) 68.06% | vs BF16 | L78 @ main |
| drawthingsai/GGUF Size includes MTP weights; size without MTP 7.78 GiB [L79] | DT-bpw2.48-IQ2_S | GGUF | 8.8 | download | — | — | — | — | AIME 2026 100.00%; GPQA Diamond 77.27%; IFEval 93.16%; BFCL v4 (1k) 69.56% | vs BF16 | L79 @ main |
| drawthingsai/GGUF Vision encoder paired with DT-IQ3_XXS language model, 100-sample subsets; size 888.01 MiB (MiB not a supported unit); MSE/cosine not mapped [L93,L106] | mmproj BF16 (vision encoder) | GGUF | — | — | — | — | — | RealWorldQA 73%; OCRBench 86% | vs BF16 | L93 @ main | |
| drawthingsai/GGUF 100-sample subsets; size in MiB [L94,L106] | mmproj FP16 (vision encoder) | GGUF | — | — | — | — | — | RealWorldQA 73%; OCRBench 86% | vs BF16 | L94 @ main | |
| drawthingsai/GGUF 100-sample subsets; size in MiB [L95,L106] | mmproj Q8_0 (vision encoder) | GGUF | — | — | — | — | — | RealWorldQA 71%; OCRBench 86% | vs BF16 | L95 @ main | |
| drawthingsai/GGUF 100-sample subsets; size in MiB [L96,L106] | mmproj Q6_K (vision encoder) | GGUF | — | — | — | — | — | RealWorldQA 73%; OCRBench 86% | vs BF16 | L96 @ main | |
| drawthingsai/GGUF 100-sample subsets; size in MiB [L97,L106] | mmproj Q5_K (vision encoder) | GGUF | — | — | — | — | — | RealWorldQA 74%; OCRBench 86% | vs BF16 | L97 @ main | |
| drawthingsai/GGUF 100-sample subsets; size in MiB [L98,L106] | mmproj Q4_K (vision encoder) | GGUF | — | — | — | — | — | RealWorldQA 71%; OCRBench 85% | vs BF16 | L98 @ main | |
| DuoNeural/TAP-DPQ-v2-GGUF PPL lower than BF16 is claimed [L45,L154]; small N benchmarks | Qwen3.8-27B-TAP-DPQ-v2-Q4_K_M.gguf | GGUF | 16.81 | download | — | — | 3.6239 🔥 | — | GSM8K 87.0% (87/100); HumanEval 60.0% (30/50); ToolCalling 93.3% (14/15) | vs BF16 | L58 @ main |
| DuoNeural/TAP-DPQ-v2-GGUF | Qwen3.8-27B-TAP-DPQ-v2-IQ3_XXS.gguf | GGUF | 11.96 | download | — | — | 4.0114 | — | GSM8K 90.0% (90/100); HumanEval 56.0% (28/50); ToolCalling 93.3% (14/15) | vs BF16 | L59 @ main |
| DuoNeural/TAP-DPQ-v2-GGUF | Qwen3.8-27B-TAP-DPQ-v2-IQ2_M.gguf | GGUF | 10.78 | download | — | — | 6.3353 | — | GSM8K 88.0% (88/100); HumanEval 54.0% (27/50); ToolCalling 93.3% (14/15) | vs BF16 | L60 @ main |
| DuoNeural/TAP-DPQ-v2-GGUF Card lists IQ1_S at ~2.21 bpw [L62] which is above IQ2_XXS ~2.06 | Qwen3.8-27B-TAP-DPQ-v2-IQ1_S.gguf | GGUF | 7.6 | download | — | — | 6.0671 ⚡ | — | GSM8K 45.0% (45/100); HumanEval 18.0% (9/50); ToolCalling 20.0% (3/15) | vs BF16 | L62 @ main |
| eaddario/GGUF top1_pct from 'Same Top-P' column [L118]benchmark table also lists F16 only via PPL row; speed table (L139-164) omitted | Qwen3.8-27B-Q8_0 | GGUF | 27.00 | download | 0.001442 ±0.000061 | 98.134 ±0.035 | 6.671738 ±0.042653 | — | ARC Challenge 64.0000 ±1.7539; GPQA-Diamond 32.3232 ±3.3323; HellaSwag 82.40; MMLU-Redux 82.2667 ±1.3956; Truthful QA 33.4667 ±1.7242; WinoGrande 75.7333 ±1.5664; Avg Score 61.70 | other or no reference | L118 @ main |
| eaddario/GGUF top1_pct from 'Same Top-P' column [L117] | Qwen3.8-27B-Q7_K | GGUF | 24.00 | download | 0.002050 ±0.000086 | 97.913 ±0.037 | 6.678988 ±0.042743 | — | ARC Challenge 64.2667 ±1.7510; GPQA-Diamond 32.8283 ±3.3457; HellaSwag 82.27; MMLU-Redux 82.2667 ±1.3956; Truthful QA 33.7333 ±1.7276; WinoGrande 76.0000 ±1.5605; Avg Score 61.89 | other or no reference | L117 @ main |
| eaddario/GGUF top1_pct from 'Same Top-P' column [L116] | Qwen3.8-27B-Q6_K | GGUF | 21.00 | download | 0.004195 ±0.000167 | 97.225 ±0.043 | 6.661399 ±0.042549 | — | ARC Challenge 63.7333 ±1.7567; GPQA-Diamond 33.8384 ±3.3711; HellaSwag 82.40; MMLU-Redux 81.8667 ±1.4078; Truthful QA 34.0000 ±1.7309; WinoGrande 76.0000 ±1.5605; Avg Score 61.97 | other or no reference | L116 @ main |
| eaddario/GGUF top1_pct from 'Same Top-P' column [L115] | Qwen3.8-27B-Q5_K | GGUF | 18.00 | download | 0.009866 ±0.000216 | 95.745 ±0.052 | 6.696776 ±0.042814 | — | ARC Challenge 64.0000 ±1.7539; GPQA-Diamond 31.3131 ±3.3042; HellaSwag 82.40; MMLU-Redux 82.0000 ±1.4038; Truthful QA 32.9333 ±1.7172; WinoGrande 76.1333 ±1.5576; Avg Score 61.46 | other or no reference | L115 @ main |
| eaddario/GGUF top1_pct from 'Same Top-P' column [L114] | Qwen3.8-27B-Q4_K | GGUF | 14.00 | download | 0.022733 ±0.000324 | 93.505 ±0.064 | 6.736913 ±0.043056 | — | ARC Challenge 63.4667 ±1.7594; GPQA-Diamond 32.3232 ±3.3323; HellaSwag 82.00; MMLU-Redux 81.3333 ±1.4237; Truthful QA 34.4000 ±1.7358; WinoGrande 75.4667 ±1.5722; Avg Score 61.50 | other or no reference | L114 @ main |
| eaddario/GGUF top1_pct from 'Same Top-P' column [L113] | Qwen3.8-27B-Q3_K | GGUF | 11.00 | download | 0.082257 ±0.000671 | 87.799 ±0.085 | 6.960375 ±0.043746 | — | ARC Challenge 62.9333 ±1.7648; GPQA-Diamond 34.3434 ±3.3832; HellaSwag 81.20; MMLU-Redux 82.6667 ±1.3831; Truthful QA 34.6667 ±1.7389; WinoGrande 74.8000 ±1.5864; Avg Score 61.77 | other or no reference | L113 @ main |
| eaddario/GGUF top1_pct from 'Same Top-P' column [L112] | Qwen3.8-27B-Q2_K | GGUF | 8 | download | 0.288783 ±0.001496 | 77.247 ±0.109 | 8.213575 ±0.052813 | — | ARC Challenge 66.2667 ±1.7276; GPQA-Diamond 30.3030 ±3.2743; HellaSwag 76.00; MMLU-Redux 78.1333 ±1.5103; Truthful QA 32.1333 ±1.7063; WinoGrande 71.8667 ±1.6430; Avg Score 59.12 | other or no reference | L112 @ main |
| eaddario/GGUF top1_pct from 'Same Top-P' column [L111]rhoPPL (c8) is a correlation, omitted | Qwen3.8-27B-Q1_L | GGUF | 5.6 | download | 0.931233 ±0.002887 | 60.773 ±0.127 | 15.076645 ±0.103470 | — | ARC Challenge 52.5333 ±1.8246; GPQA-Diamond 23.7374 ±3.0314; HellaSwag 62.53; MMLU-Redux 54.8000 ±1.8185; Truthful QA 30.0000 ±1.6744; WinoGrande 63.2000 ±1.7621; Avg Score 47.80 | other or no reference | L111 @ main |
| EliovpAI/W3Rot-INT3-Paiton-RDNA4 size 9.55 GB from header line [L27], not the same tableMMLU-Pro log-likelihood subset [L130]GSM8K/HumanEval/needle not re-run for these weights [L148] | These weights (12 Oct 2026 scale-distilled) | other | 9.6 | download | — | — | — | — | kld:decode 0.0673; kld:prefill 0.0818; top1:decode 89.70; top1:prefill 88.38; kld:long_decode 0.0746; kld:long_prefill 0.0909; kld:r9700_served 0.1395; top1:r9700_served 88.14 % | vs BF16 | L27 @ main |
| empero-ai/Ridge-GGUF PPL 7.82 +/- 0.14 [L111]card also lists 12.59 GB size [L59] | Ridge-3.7bpw | GGUF | 12.59 | download | — | — | 7.82 ± 0.14 | +9.3 % | — | vs BF16 | L111 @ main |
| EschaLabs/Escha-W2 2-bit quant; Commonsense-6 ran on L40, GPQA/LCB on RTX 4090 vs FP8 on different box [L482-485]Benchmarks measured with SPEC=0 [L298] | Escha-W2 | other | 10.15 | download | — | — | — | — | boolq 88.20; hellaswag 80.19; piqa 81.18; winogrande 74.98; GPQA-Diamond 88.38; LiveCodeBench v6 86.81 | other or no reference | L435 @ main |
| fasolack/W4A16 Size stated in prose as about 19.5 GB vs about 52 GB BF16 [L35]Own BF16 baseline was not measured [L230-231]HumanEval gap to reference partly sampling noise, rest likely calibration mix [L246-258]Table values include +/- confidence half-widths [L276-279] | This build (int4 W4A16) | INT4 | 19.50 | download | — | — | — | — | GSM8K 97.04% ± 0.47%; MMLU-Pro 81.36% ± 1.01% | other or no reference | L35 @ main |
| FedericoSciuca/depthaware-GGUF excess over reference +0.4640 c6 is absolute PPL difference, not percent [L36]perplexity is a proxy, not a task score [L80] | this model, measured band 51-63 (depth-aware 2-bit) | GGUF | 12.48 | download | — | — | 6.0590 | — | — | vs a Q8_0 | L36 @ main |
| finalvent/OddEye-0.1-27B-NVFP4 task-specific fine-tune (self-distillation) on Qwen3.8-27B, not a plain quant [L121-122]LLM-graded, read as relative comparison not absolute accuracy [L94]images/hour on RTX PRO 6000, 32 concurrent [L88] | OddEye | NVFP4 | 20.60 | download | — | — | — | — | ReferenceFactsCaptured 0.53; ReferenceFactsContradictedPct 11.3%; OverconfidentPerPhoto 0.69; ClearErrorsPerPhoto 0.26 | vs BF16 | L77 @ main |
| fraQtl/Hi-Fi-GGUF KLD cells include percent change vs canonical in parentheses [L34]task benchmark differences are statistical ties per card [L59-60] | fraQtl Hi-Fi (this repo) | GGUF | 16.46 | download | — | — | — | — | kld:code/math 0.12297 (−45.3%); kld:general 0.23120 (−18.0%); GSM8K 95.0% (190/200); MATH-500 87.6% (438/500) | vs a Q8_0 | L34 @ main |
| Freaksterz/SmoothQuant-W8A8-INT8 floor-free KLD 0.00371 (c7), PPL delta-ln +0.0030 (c9) and DFlash2 acceptance length 4.34 (c10) also in table [L29]size from prose, 29 GB on disk [L103] | v3 W8A8 (main) | INT8 | 29.00 | download | 0.00556 | 96.9% | — | — | — | vs BF16 | L29 @ main |
| Frosty40/GGUF size shown is the -MTP file; non-MTP file listed separately [L28-L29]; backbone tensors byte-identical [L57]tail is the 99.9th-percentile per-token KLD, mapped to kld_99 key [L47-L48]top1 is greedy same-top agreement with bf16 [L48] | Q6_K-QKV8(-MTP) | GGUF | 23.20 | download | 0.00163 kld:prose | — | — | — | top1:prose 97.75; kld:code 0.00084; top1:code 99.80; kld:agentic 0.00094; top1:agentic 99.27 | vs BF16 | L28 @ main |
| GestaltLabs/EXL3-11.5GB KL measured against FP8, not BF16 [L122]domain KL values live in prose lines; ids c50-c53 point at those cells [L133-136]task accuracy is a pass count on 128 tasks, lower than prior release [L146-148] | This Release (Expanded 8K, 598 rows) | EXL3 | 11.50 | download | — | — | — | — | kld:development 0.16220; kld:acceptance 0.25899; kld:combined 0.21060 | other or no reference | L128 @ main |
| glyd/kestrel card says about 6.5 bits per weight in prose, no table cell [L15]KL measured on WikiText-2 [L32] | kestrel (this repo) | other | 22.00 | download | 0.00477 | — | — | — | GSM8K 97.3% | vs BF16 | L29 @ main |
| glyd/penguin Lossless: every weight bit-identical to bf16 [L15]; about 11.0 bits per weight stated at L43 | penguin (lossless, this repo) | other | 37.00 | download | 0 | — | — | — | — | vs BF16 | L33 @ main |
| glyd/swift about 5.5 bits per weight in prose [L15]text only, vision excluded [L57] | swift (this repo) | other | 18.60 | download | 0.0139 | — | — | — | GSM8K 97.2% | vs BF16 | L30 @ main |
| gratex/W4A16-g128-sym-GPTQ KLD uses sparse top-20 logprobs and underestimates absolute KLD by about 10-15% vs full vocab [L116]kld_99 is the p99 KLD row [L122]BBH reported value re-scored with markdown stripping; raw harness score differs [L155]HumanEval variant extracts fenced blocks; stricter extraction scores lower [L157]downstream scores are fractions, no BF16 comparison column [L136-L153] | This model (v9-2 g128 sym) | GPTQ | 19.00 | download | 0.0699 | 92.61% | 9.8108 | +1.77% | kld_99 0.5619; MMLU 0.839; MMLU-Pro 0.823; Global-MMLU-en 0.868; BBH 0.924 ¹; Winogrande 0.795; TruthfulQA-mc1 0.361; TruthfulQA-mc2 0.544; IFEval-strict 0.887; IFEval-loose 0.915; HumanEval 0.976; MBPP 0.874; BABILong 0.876; LongBench 0.520 | vs BF16 | L110 @ main |
| greghavens/bnb-4bit Perplexity cell holds 6.5571 vs 6.4833 bf16 (+1.14%) in one candidate, same id used for ppl and delta [L54, L176]IFEval prompt-level strict NF4 cell has no candidate id (delta +0.00pp vs bf16 0.8189) [L192]BFCL live_irrelevance alone drops -3.39pp [L292]Over-calling: spurious tool call on 27.0% of no-op prompts vs base 23.8% [L45]Tool-call prompts showed mean KL 0.93 and 12% high-confidence top-1 flips [L464] | NF4 (this repo, bf16 vision tower) | bitsandbytes | 18.58 | resident | — | — | — | — | ppl:wikitext-2 6.5571 vs 6.4833 bf16 (+1.14%); IFEval-prompt-loose 0.8447; BFCL-overall 0.8113; NoveltyBench-distinct_10 4.99; NoveltyBench-utility_10 5.359 | vs BF16 | L54 @ main |
| gubernac/unc-fp8 Deltas reflect the abliteration edit, since FP8 scheme matches official [L214-215]Refusal tables are abliteration effects, omitted [L164-200]Weights ~31 GB [L312] | This model (abliterated block-FP8) | FP8 | 31.00 | download | — | — | — | — | MMLU 84.7%; MMLU-Pro 76.8%; GSM8K 88.7%; CMMLU 80.8% | other or no reference | L208 @ main |
| guideboardlabs/Q3-DOWN-XS-GGUF Quality % is relative to UD-IQ3_XXS PPL, not bf16; omitted [L32]prose gives 7.91 GiB vs table 7.68 GiB [L22,L30] | Qwen3.8-27B-Q3-DOWN-XS.gguf | GGUF | 8.2 | download | — | — | — | — | ppl:wikitext-2 8.2691 | other or no reference | L30 @ main |
| HangGlidersRule/Darkstar-Base-ModelOpt-W4A16-NVFP4-Mixed-FP8 GPQA 153/198, thinking off [L43]refusal cells are fraction and percent in one cell [L45-L46] | Darkstar W4A16-NVFP4-Mixed-FP8 | NVFP4 | — | — | — | — | — | GPQA-Diamond 77.27 | vs BF16 | L43 @ main | |
| Hent-AI/EXL3-1.14514bpw 7.83 GB weight-file size stated in prose only [L54]bpw is decoder Linear target-weight average [L206]no reference-model comparison; Japanese eval is custom, not JGLUE reproduction [L118] | Qwen3.8-27B EXL3 1.14514bpw | EXL3 | — | — | — | — | — | MMLU-Redux 31.30%; MuSR 33.99%; GSM8K 16.60%; HumanEval+ 31.71%; IFEval 60.78%; BFCL v3 62.43%; JCommonsenseQA thinking-off 73.01%; JCommonsenseQA thinking-on 77.57%; JSQuAD EM thinking-off 85.37%; JSQuAD EM thinking-on 86.88%; JSQuAD F1 thinking-off 93.04%; JSQuAD F1 thinking-on 94.32%; MGSM-ja thinking-off 18.80%; MGSM-ja thinking-on 18.40% | reference not stated | L69 @ main | |
| HermiHg/mmproj-Q5_K-MIX-GGUF mmproj projector only, scores are accuracy fractions [L46-57]size 331.1 MiB (5.99 bpw) vs 888.0 MiB BF16 printed in row label L58, no candidate id; ratio c84 = 0.37BLINK counting and object localization Q5_K-MIX cells absent from candidates; card says exactly matched [L55-56]Q8_0 / Q4_K failing-document counts (L78-84) are a different, selection-biased calibration set; not mapped | Q5_K-MIX | GGUF | — | — | — | — | — | TextVQA 0.8144; OCRBench-v2 0.6613; GQA 0.6240; ScienceQA 0.9475; MMStar 0.7113; ChartQA 0.8460; BLINK jigsaw 0.5067; BLINK relative depth 0.8952; BLINK visual similarity 0.8963; Overall 0.7458 | vs BF16 | L46 @ main | |
| HivenetQuant/NVFP4 niah NVFP4 cell not a candidate; only delta c26 given [L77]Delta columns are as printed (points) | NVFP4 (FP8-protected release) | NVFP4 | — | — | — | — | — | mmlu_pro 77.26; aime25 96.00; ifeval 90.94; bfcl 79.18 | other or no reference | L69 @ main | |
| huginnfork/FP8 192 modules quantised [L70]Negative dPPL does not mean it beats its parent, within noise [L80-81] | This model (huginnfork FP8) | FP8 | — | — | — | 6.9325 | -0.13 % | kld:weight_only 0.0362; kld:as_deployed 0.0756 | vs BF16 | L70 @ main | |
| huginnfork/NVFP4A16 KLD is the weight-only column; no as-deployed value given [L70]192 modules quantised [L70] | This model (NVFP4A16, MLPs only) | NVFP4 | — | 0.1267 | — | 7.1497 | +3.00 % | — | vs BF16 | L70 @ main | |
| ilintar/gguf-strix-halo Perplexity 15.3977 +/- 0.73357 on 16 held-out 512-token chunks [L40]Quant type of main model not named in provided linesDFlash2 draft IQ4_XS vs Q8_0 speed/acceptance tables omitted (speed only) [L58-72] | Main-model quantization | GGUF | — | — | — | perplexity was 15.3977 +/- 0.73357, comp | — | — | vs BF16 | L40 @ main | |
| inferencerlabs/MLX-Q9 Token Accuracy = percent correctly generated base tokens [L46]bpw taken from row label 'Q9' [L39] | Q9 | MLX | — | — | 97.80% | 1.20312 | — | — | vs BF16 | L39 @ main | |
| intel-ai/MXFP8-CT-AutoRound size is weights loaded in vLLM on B300, from prose [L134]effective bits/param cell holds two values (weights only vs including others), omitted [L42] | MXFP8 (this checkpoint) | MXFP8 | 30.44 | GPU weights | — | — | — | — | gsm8k_strict 0.9719; mmlu 0.8348; piqa_acc 0.8085; piqa_acc_norm 0.8156; hellaswag_acc 0.6361; hellaswag_acc_norm 0.8283; AVG 0.8128 | vs BF16 | L87 @ main |
| intel-ai/MXFP8-MXFP8MTP-CT-AutoRound gsm8k is strict-match; others 0-shot acc [L109-115]acceptance rate 0.9451 is a speed-side metric, not mapped [L132] | MXFP8 + MTP-MXFP8 (this checkpoint) | MXFP8 | — | — | — | — | — | mmlu 0.8331; piqa_acc_norm 0.8139; hellaswag_acc_norm 0.8272 | vs BF16 | L96 @ main | |
| ionos/GPTQ-Int4 MMMU checkpoint cell missing from table (only base 81.0% present) [L44] | This checkpoint (GPTQ-Int4) | GPTQ | — | — | — | — | — | OmniDocBench_reading_order 92.8%; OmniDocBench_display_formula 72.9%; MathVista 88.3%; V*Bench 87.0%; MMLU-Pro 85.5%; MMLU-DE 79.5%; BFCL 58.0%; IFEval 38.4% | vs BF16 | L40 @ main | |
| ISTA-DASLab/3Bit-GSQ AIME 2025 3-bit GSQ cell not in candidates [L44]size 11.83 GB with vision, 10.90 GB text-only [L115,L135]3-bit GSQ group 128, embed/head 4-bit RTN group 64 [L26-28] | 3-bit GSQ | other | 11.83 | download | — | — | — | — | GPQA Diamond 91.41 | other or no reference | L45 @ main |
| IsValorum/VAL-APEX-I-MiniPlus-V3-GGUF PPL 6.1361 +/- 0.4885 [L73]; delta cell combines absolute +0.3419 and +5.90%NanoBench agentic harness scores (L86-90) were run by a community member and linked to a different repo (Huihui NanoPlus), not mapped | Qwen3.8-27B-VAL-APEX-I-MiniPlus-V3.gguf | GGUF | 14.16 | download | — | — | 6.1361 ± 0.4885 | — | — | vs BF16 | L73 @ main |
| IsValorum/VAL-APEX-I-NanoPlus-V3-GGUF whole-file BPW incl. metadata and MTP head [L151]PPL +/- 0.5011 stated in cell [L73] | Qwen3.8 27B Base - NanoPlus V3 (this release) | GGUF | 12.18 | download | — | — | — | — | ppl:wikitext2 6.2577 ± 0.5011 | vs BF16 | L73 @ main |
| jaakkomoller/q3mix-NInfer MTP3 acceptance 86.67% in front-matter metadata [L37]; pass counts out of 175 and 35 | q3mix | other | — | — | — | — | — | LiveCodeBench_v6_175 43/175; quality_gate_35 29/35 | other or no reference | L179 @ main | |
| Jackrong/Qwopus3.8-27B-Flash-GGUF Fine-tune of Qwen3.8-27B, not a plain quant; author-provided local results [L143]MMLU-Pro 2,500-question mixed set accuracy [L157]Agentic battery 13/14 passed on one RTX 5090 [L162]; speed/MTP acceptance figures omitted | Qwopus3.8-27B-Flash (Q5_K_M + MTP) | GGUF | — | — | — | — | — | MMLU-Pro 91.28% | other or no reference | L157 @ main | |
| JANGQ-AI/JANGH2 All KL values are median KL vs bf16, not mean [L50-55]Size 14.25 GiB stated in prose headline [L39] | JANGH2 | other | 15.30 | download | — | — | — | — | kld:mixed_held_out 0.037; top1:mixed_held_out 80.1 %; kld:agentic_tool_conversations 0.061; top1:agentic_tool_conversations 75.3 %; kld:reasoning 0.008; top1:reasoning 85.4 %; kld:own_traces_low 0.008; kld:own_traces_medium 0.002; kld:own_traces_xhigh 0.026; top1:own_traces_low 81.2; top1:own_traces_medium 88.1; top1:own_traces_xhigh 81.9 | vs BF16 | L39 @ main |
| JANGQ-AI/JANGH2 Median KL per topic [L64-79] | JANGH2 by topic (held-out text) | other | — | — | — | — | — | kld:coding 0.0092; top1:coding 89.5 %; kld:math 0.0068; top1:math 90.8 %; kld:office_finance 0.0087; top1:office_finance 86.9 %; kld:topic_reasoning 0.0081; top1:topic_reasoning 81.0 %; kld:science 0.0460; top1:science 87.1 %; kld:long_context 0.0446; top1:long_context 80.4 %; kld:cybersecurity 0.0368; top1:cybersecurity 80.5 %; kld:networking 0.0244; top1:networking 81.9 %; kld:terminal_shell 0.0738; top1:terminal_shell 73.5 %; kld:agentic_tool_calls 0.0975; top1:agentic_tool_calls 71.2 %; kld:creative_writing 0.1989; top1:creative_writing 74.2 %; kld:multi_turn_agentic 0.0615; top1:multi_turn_agentic 75.3 % | vs BF16 | L68 @ main | |
| JANGQ-AI/JANGH2 Median KL per topic on own reasoning traces [L81-91] | JANGH2 by topic (own bf16 reasoning traces) | other | — | — | — | — | — | kld:traces_coding 0.00508; top1:traces_coding 89.4 %; kld:traces_math 0.00061; top1:traces_math 91.6 %; kld:traces_office_finance 0.00007; top1:traces_office_finance 91.2 %; kld:traces_science 0.00779; top1:traces_science 87.8 %; kld:traces_networking 0.00009; top1:traces_networking 92.8 %; kld:traces_cybersecurity 0.08516; top1:traces_cybersecurity 81.1 %; kld:traces_agentic_tool_calls 0.03968; top1:traces_agentic_tool_calls 72.6 % | vs BF16 | L85 @ main | |
| JANGQ-AI/JANG_4D KL is median KL, not mean [L50]Size 21 GiB stated in prose only [L38]Think-block close decisions agree 40/40 and 65/65; tool-call vs answer decisions flip on 9/263 agentic positions [L59-60] | JANG_4D (quality vs bf16, by set) | other | — | — | — | — | — | kld:mixed_held-out_text 0.0007; top1:mixed_held-out_text 95.2 %; kld:agentic_and_tool_conversations 0.0014; top1:agentic_and_tool_conversations 92.1 %; kld:reasoning 0.0001; top1:reasoning 96.9 %; kld:bf16_traces_effort_low 0.0001; top1:bf16_traces_effort_low 93.9 %; kld:bf16_traces_effort_medium 0.00004; top1:bf16_traces_effort_medium 97.7 %; kld:bf16_traces_effort_xhigh 0.0005; top1:bf16_traces_effort_xhigh 95.1 % | vs BF16 | L52 @ main | |
| JANGQ-AI/JANG_4D Median KL per topic, same rows as JANGH2 [L66]; tokens scored per topic in table (e.g. coding c14, multi-turn agentic c69) | JANG_4D (by topic, held-out text set) | other | — | — | — | — | — | kld:coding 0.0002; top1:coding 98.4 %; kld:math 0.0001; top1:math 98.7 %; kld:office_finance 0.0001; top1:office_finance 97.7 %; kld:reasoning 0.0001; top1:reasoning 94.5 %; kld:science 0.0010; top1:science 98.0 %; kld:long_context_attention_recall 0.0007; top1:long_context_attention_recall 96.6 %; kld:cybersecurity 0.0008; top1:cybersecurity 94.9 %; kld:networking 0.0011; top1:networking 93.7 %; kld:terminal_shell 0.0013; top1:terminal_shell 92.5 %; kld:agentic_tool_calls 0.0026; top1:agentic_tool_calls 89.8 %; kld:creative_writing 0.0028; top1:creative_writing 96.5 %; kld:multi_turn_agentic 0.0014; top1:multi_turn_agentic 92.1 % | vs BF16 | L70 @ main | |
| JANGQ-AI/JANG_4D Thinking + answer tokens of own bf16 traces [L83]; median KL | JANG_4D (own bf16 reasoning traces, by topic) | other | — | — | — | — | — | kld:coding 0.00007; top1:coding 98.4 %; kld:math 0.00001; top1:math 98.6 %; kld:office_finance 0.00000; top1:office_finance 98.5 %; kld:science 0.00010; top1:science 98.1 %; kld:networking 0.00000; top1:networking 98.7 %; kld:cybersecurity 0.00135; top1:cybersecurity 97.3 %; kld:agentic_tool_calls 0.00093; top1:agentic_tool_calls 89.7 % | vs BF16 | L87 @ main | |
| jaromer/0bserverx-Heretic-Abliterated-Uncensored-GGUF KL is abliteration damage vs base, not a quantization KLD; file/quant it was measured on not stated [L126-130]refusals 0-1/100 (c7.0, c7.1) are not a quant-quality metric key | RVN (this repo) | GGUF | — | 0.0085 | — | — | — | — | other or no reference | L130 @ main | |
| jcbtc/CIRU-ActiveFPX-PromptForge GGUF alone 15.44 GiB, full payload 35.81 GiB, not in metric table [L102]same-top token agreement 91.152% [L57] | V3 | other | — | 0.031457 | 91.152% | — | 1.016361 | — | vs BF16 | L55 @ main | |
| jcbtc/IU4-Kairic-Edge Configured-system comparison, not a one-variable quantization experiment [L115]Run predates v1.2 correction [L117]GGUF is 15.48 GiB (16,617,792,672 bytes) in separate files table [L225]; total with companions 26.05 GiB [L229] | Kairic Edge IU4 v1/v1.1 | other | — | — | — | — | — | HumanEval 158/164 (96.34%); HumanEval+ 152/164 (92.68%) | reference not stated | L104 @ main | |
| jclyons52/Unsloth-Dynamic-v3-mlx ppl includes +/- 0.212 error [L46]peak memory 13.8 GB, decode 10.17 t/s [L47-48] | Unsloth Dynamic v3.0 allocation, native MLX | MLX | 12.40 | download | — | — | — | — | ppl:wikitext-2 Perplexity: 12.724 ± 0.212 | reference not stated | L21 @ main |
| jdominick05/24GB-262k-GGUF size column is MiB [L40]; file table lists 17.0 GiB [L28]median KLD also reported [L42]; PPL is relative to Q8_0 (+0.1%)99.9th pct KLD mapped to kld_99 [L42] | R3-mix (Qwen3.8-27B-R3-mix.gguf) | GGUF | 18.23 | download | 0.0142 ± 0.0011 | 96.10% | — | +0.1% | — | vs a Q8_0 | L42 @ main |
| jdominick05/24GB-262k-GGUF measured on same weights without the MTP layer [L49]size column is MiB [L40]; file table lists 16.4 GiB [L29]99.9th pct KLD mapped to kld_99 [L43] | R5-mix-MTP | GGUF | 17.65 | download | 0.0165 ± 0.0014 | 95.90% | — | +0.3% | — | vs a Q8_0 | L43 @ main |
| jialinyyzz/abliterated-MLX-4bit 4.695 bits/weight stated in prose [L32]capability and refusal numbers measured on bf16 master; 4-bit quant error not measured [L65-66,L78-79]~19 GB peak RAM [L33] | abliterated MLX 4-bit (w177) | MLX | 15.00 | download | — | — | — | — | bpw 4.695 bits/weight | vs BF16 | L32 @ main |
| jialinyyzz/abliterated-MLX-8bit KL (orig || abliterated) measures the abliteration edit on harmless prompts, 8-bit MLX affine group size 64 [L32, L44]HellaSwag w177 value not in candidates (only original and delta 0.000)Refusal counts (base 98, w177 11 of 100) are not a mapped quality metric [L55-56] | abliterated MLX 8-bit (w177) | MLX | — | 0.056 | — | — | — | ARC-Challenge 0.567; Winogrande 0.770; OpenBookQA 0.470; GSM8K 0.78 | other or no reference | L44 @ main | |
| Jianqiao1/MoQ-GGUF [L54] | Qwen3.8-27B-MTP-MoQ-5.1.gguf | GGUF | 17.34 | download | 0.012857 | 94.968% | 7.020549 | — | kld_99 0.123491 | vs BF16 | L54 @ main |
| Jianqiao1/MoQ-GGUF [L53] | Qwen3.8-27B-MTP-MoQ-4.9.gguf | GGUF | 16.41 | download | 0.016342 | 94.423% | 7.021034 | — | kld_99 0.162814 | vs BF16 | L53 @ main |
| Jianqiao1/MoQ-GGUF [L52] | Qwen3.8-27B-MTP-MoQ-4.8.gguf | GGUF | 16.04 | download | 0.017282 | 94.339% | 7.008832 | — | kld_99 0.169743 | vs BF16 | L52 @ main |
| Jianqiao1/MoQ-GGUF [L51] | Qwen3.8-27B-MTP-MoQ-4.6.gguf | GGUF | 15.13 | download | 0.020171 | 93.809% | 7.014436 | — | kld_99 0.200817 | vs BF16 | L51 @ main |
| Jianqiao1/MoQ-GGUF [L50] | Qwen3.8-27B-MTP-MoQ-4.3.gguf | GGUF | 14.89 | download | 0.024500 | 92.648% | 7.045628 | — | kld_99 0.216612 | vs BF16 | L50 @ main |
| Jianqiao1/MoQ-GGUF [L49] | Qwen3.8-27B-MTP-MoQ-4.1.gguf | GGUF | 14.14 | download | 0.031460 | 91.872% | 7.068801 | — | kld_99 0.285069 | vs BF16 | L49 @ main |
| Jianqiao1/MoQ-GGUF [L48] | Qwen3.8-27B-MTP-MoQ-3.8.gguf | GGUF | 12.71 | download | 0.048190 | 90.109% | 7.122706 | — | kld_99 0.437946 | vs BF16 | L48 @ main |
| Jianqiao1/MoQ-GGUF [L47] | Qwen3.8-27B-MTP-MoQ-3.6.gguf | GGUF | 11.89 | download | 0.072244 | 88.722% | 7.221080 | — | kld_99 0.717370 | vs BF16 | L47 @ main |
| Jianqiao1/MoQ-GGUF Same top-p mapped to top1_pct [L44][L46] | Qwen3.8-27B-MTP-MoQ-3.2.gguf | GGUF | 10.57 | download | 0.111637 | 85.679% | 7.408463 | — | kld_99 1.084846 | vs BF16 | L46 @ main |
| jkim96/DASHQ-Q2-GGUF [L43] | DASH-Q Q2_K_XL | GGUF | 9.8 | download | — | — | — | — | ppl:wikitext2 6.40; ppl:c4 10.92 | other or no reference | L27 @ main |
| jkim96/DASHQ-Q2-GGUF [L41] | DASH-Q IQ2_S | GGUF | 8.7 | download | — | — | — | — | ppl:wikitext2 6.72; ppl:c4 11.51 | other or no reference | L26 @ main |
| jkim96/DASHQ-Q2-GGUF [L39] | DASH-Q IQ2_XS | GGUF | 8.6 | download | — | — | — | — | ppl:wikitext2 6.75; ppl:c4 11.61 | other or no reference | L25 @ main |
| jkim96/DASHQ-Q2-GGUF [L37] | DASH-Q IQ2_XXS | GGUF | 7.8 | download | — | — | — | — | ppl:wikitext2 7.62; ppl:c4 13.39 | other or no reference | L24 @ main |
| Jon-Nielsen/exl3-4.00bpw Size is VRAM peak ~24 GB with Q6 KV 262K config, not file size; file ~16 GB in prose [L11, L20]Pre-release benchmarks [L14] | EXL3 4.00bpw | EXL3 | 24.00 | GPU weights | — | — | — | — | ppl:40K_tech 5.25 | reference not stated | L20 @ main |
| Jon-Nielsen/exl3-6.00bpw Pre-release benchmarks; size is VRAM peak at runtime [L14,L21]Upstream Qwen3.8-27B tables pasted below are not measurements of this quant [L80-244] | 6.00bpw | EXL3 | 30.00 | GPU weights | — | — | — | — | ppl:40K_tech 5.07 | reference not stated | L21 @ main |
| JRamirez-UAB/EXL3-3.85bpw-24GB internal and public (turboderp trace) values from same row [L89]BF16 ppl 1.5860 internal / 1.3332 public [L85] | EXL3 3.85 bpw (main) | EXL3 | 14.62 | download | 0.0073 | — | 1.5976 | — | kld:public_trace 0.0067; ppl:public_trace 1.3430 | vs BF16 | L89 @ main |
| JRamirez-UAB/EXL3-3.85bpw-24GB [L90] | EXL3 3.70 bpw (3.70bpw branch) | EXL3 | 14.16 | download | 0.0089 | — | 1.5999 | — | kld:public_trace 0.0085; ppl:public_trace 1.3466 | vs BF16 | L90 @ main |
| JRamirez-UAB/GPTQ-W4A16-embed-int4-24GB Weights column of family table [L91]; mean KLD, medians also reported (internal 0.0127, public 0.0017) [L91]Agent tasks vary about +-2 of 30 between runs [L105]Card states quality not evaluated with standard benchmarks [L452] | GPTQ W4A16 + INT4 embeddings (mtp-int8 / main) | GPTQ | 15.26 | download | — | — | — | — | kld:internal 0.0400; kld:public 0.0312; ppl:internal 1.6631; ppl:public 1.3779 | vs BF16 | L91 @ main |
| JRamirez-UAB/GPTQ-W4A16-g64-embed-int8-24GB median KLD also reported [L83]GPTQ evaluated as BF16 copy of dequantized weights [L95] | GPTQ W4A16 g64 + INT8 embeddings (v4) | GPTQ | 16.91 | download | — | — | — | — | kld:internal 0.0175; kld:public 0.0171; ppl:internal 1.6203; ppl:public 1.3568 | vs BF16 | L83 @ main |
| julianmb/Qwen-3.8-27B-ROCmFP4-FAST-GGUF PPL cell includes a ± error term [L188]card also claims ~99% benchmark retention without a table [L92] | ROCmFP4_FAST (this repo, original build) | GGUF | 14.55 | download | — | — | — | — | ppl:wikitext-2 7.205 ± 0.121 | reference not stated | L78 @ main |
| julianmb/Qwen-3.8-27B-ROCmFP4-FAST-GGUF PPL cell includes a ± error term [L189]size not given for this specific build; speed stated statistically identical [L191] | ROCmFP4_FAST requantized from unsloth Q8_0 | GGUF | — | — | — | — | — | ppl:wikitext-2 7.138 ± 0.120 | reference not stated | L189 @ main | |
| juzbox/NVFP4-GDN KL row L84 is custom engine; averaging method not statedsize is checkpoint size [L35] | this model (NVFP4 GDN projections, GPTQ) | NVFP4 | 19.50 | download | 0.0445 | — | — | — | BFCL v3 all (custom engine) 89.7%; BFCL parallel (custom engine) 78.5% | vs BF16 | L35 @ main |
| kacaforyah/Opti-27B median KLD 0.043 also in table [L79]unpaired ppl vs paired ratio differ [L37,L79]MMLU-Pro scorer B this release [L104] | Opti 27B | other | 11.80 | download | 0.0935 | 87.2% | 6.487 | 1.005x | ppl_ratio:paired 1.0075 +/- 0.0028; kld_99 0.876; ppl_delta_pct:wikitext-2 +0.75% +/- 0.28%; ppl_delta_pct:C4 +4.2% +/- 1.6%; ppl_delta_pct:python-stdlib +21.8% +/- 2.0%; ppl_delta_pct:pride-and-prejudice +58.9% +/- 4.2%; MMLU-Pro:scorerB 53.9 +/- 0.5 | other or no reference | L37 @ main |
| kaushikvira/nvfp4full-dflash2-NInfer asterisk on values not explained in extracted cells [L211]family table row for the v3 twin (byte-identical weights) [L223-224]rows for swift-abliterated, swift15, thinkingcap, swift15-uncensored-ajgazin are different fine-tuned models, omitted [L212-215]cross-card rows from different sessions, small deltas indicative [L206] | nvfp4full-dflash2.v3 (same weights as this v2 container) | NVFP4 | 20.08 | GPU weights | — | — | — | — | IFBench prompt-strict 65.0*; GSM8K-200 96.5%* | other or no reference | L168 @ main |
| kaushikvira/nvfp4full-dflash2-NInfer-v3 IFBench and GSM8K cells carry an asterisk with unexplained footnote [L236]size is Weights on device from validation table [L175]needle 12/12 and tool-call 10/10 gate probes are pass counts, omitted | nvfp4full-dflash2.v3 (this file) | NVFP4 | 20.08 | GPU weights | — | — | — | — | GSM8K-200 96.5%* | reference not stated | L175 @ main |
| kaushikvira/swift15-nvfp4full-dflash2-NInfer-v3 Size cell c18 is the Swift-1.0 column of the artifact-size row; this artifact column not a separate candidate [L127]Family table repeats swift15 v3 IFBench 69.0 and GSM8K 95.0% [L184]Weights in VRAM 18.0 GiB [L108] | Swift-1.5 (this artifact) | NVFP4 | 19.78 | download | — | — | — | — | IFBench prompt-loose 72.7; IFBench instr-strict 70.4; IFBench instr-loose 73.6; GSM8K-200 95.0% (190/200) | other or no reference | L118 @ main |
| kaushikvira/swift15-uncensored-ajgazin-nvfp4full-dflash2-NInfer-v3 same-session A/B [L69-70]GSM8K -0.5 pp is n=200 noise [L70]refusal-rate rows omitted (not a quality-vs-reference metric) | this artifact | NVFP4 | — | — | — | — | — | IFBench prompt-strict 70.3 (loose 74.3, instr 70.6/74.1); GSM8K-200 94.5% (−0.5 pp, n=200 noise) | other or no reference | L69 @ main | |
| kaushikvira/thinkingcap-nvfp4full-dflash2-NInfer-v3 same-session A/B [L137-141]GSM8K delta +0.5 pp is n=200 noise [L137] | ThinkingCap (this artifact) | NVFP4 | — | — | — | — | — | GSM8K-200 95.5% (191/200); IFBench prompt-strict 68.0; IFBench prompt-loose 71.0; IFBench instr-strict 67.7; IFBench instr-loose 70.3 | other or no reference | L137 @ main | |
| kayceesamuel/NF4DQ 4.1562 bits per weight [L13]Card compares against interpolated K-quant at 4.16 bpw near 7.29 [L29-30] | NF4DQ | bitsandbytes | 14.21 | download | — | — | 7.0890 ppl:wikitext | — | — | vs BF16 | L26 @ main |
| kdqemre/RQ2_K_L-GGUF Score 98/100 on 100 problems, single run [L93] | RQ2_K_L (Mixed 2-bit) | GGUF | 12.60 | download | — | — | — | — | GSM8K-100 98 | other or no reference | L91 @ main |
| kekzle/NVFP4-vllm 51.75 GiB BF16 becomes 19.2 GiB [L21] | Qwen3.8-27B-NVFP4 | NVFP4 | 20.62 | download | — | — | — | — | ppl:ppl_corpus_45k 4.6158 | other or no reference | L21 @ main |
| kingjones777/ROCmFP4-STRIX-MTP-GGUF ppl ±0.068 [L104][L89] | ROCmFP4-STRIX | GGUF | 14.76 | download | — | — | 5.8877 | — | — | other or no reference | L89 @ main |
| kingjones777/ROCmFP4-STRIX-MTP-GGUF [L90] | ROCmFP4-STRIX_LEAN | GGUF | 14.59 | download | — | — | 5.8871 | — | — | other or no reference | L90 @ main |
| kingjones777/ROCmFP4-STRIX-MTP-GGUF [L91] | ROCmFP4-FAST | GGUF | 14.31 | download | — | — | 5.9233 | — | — | other or no reference | L91 @ main |
| kirillbilchenko/MxWave-MXFP4-FP8 Absolute PPL is specific to the character-window protocol [L143-145]PPL is 0.0969% lower than H64 per comparison column on H64 row [L138]Holdout KL difference interval crosses zero; GSM8K difference not statistically proven [L154-157,L172-176]Top-1 agreement cell for candidate not given in mapped cells [L152]About 2% larger and 2.2% slower than H64 [L40-41] | MxWave MXFP4/FP8 | MXFP4 | — | — | — | — | — | ppl:wikitext2 8.111578; kld:holdout 0.044942 | vs BF16 | L139 @ main | |
| komeijishiki/EXL3-SC-3.40bpw-H6-V6 KLD is recipe-predicted, not a measured result [L135]card states achieved_bpw = 3.40 (text) but not as a table candidate [L135]per-layer SQNR 39-43 dB claimed [L136] | Qwen3.8-27B-EXL3-SC-3.40bpw-H6-V6 | EXL3 | — | 0.011 | — | — | — | — | reference not stated | L135 @ main | |
| lackonendes/PAW-125B-FLASH-NEXT-X3 two shards listed separately (42.08 and 48.28 GiB); no total cell [L60-61]bpw applies to expert bank only, decoder at Q8_0 [L18-19]MMLU-Pro and GPQA use a rescue pass with larger token cap [L78-83]model-level KL vs BF16 parent not measured [L145-146] | PAW-125B-FLASH-NEXT-X3 q8_0 (2-shard GGUF, expert bank 2.631 bpw) | other | — | — | — | — | — | MMLU-Pro 327/500 (65.4%); GPQA 117/198 (59.1%); IFBench-64 strict 15/64 (23.4%), loose 21/64 (32.8%); HumanEval 97.0; HumanEval+ 93.9; MBPP 96.3; MBPP+ 80.7; GSM8K 97/100 (97.0%) | other or no reference | L28 @ main | |
| lackonendes/PAW-27B-GGUF ~2.17 bpw stated in prose, no candidate id [L17]peer comparisons, not retention numbers [L51] | PAW-27B | GGUF | 7.8 | download | — | — | — | — | MMLU-Pro 58.6%; GSM8K 94%; HumanEval+ 87.2%; MBPP+ 75.9% | other or no reference | L41 @ main |
| lackonendes/PAW-27B-X3.1-GGUF Cells include counts and pending/running notes; GPQA 11 items pending, MMLU-Pro 20 pending, LiveCodeBench still running, IFBench 6 pending [L63-66,L141-142]Retention ratios vs published BF16 given in table [L62-66]Only 3.5 bpw file has benchmark results [L44-45,L140] | PAW-27B-X3.1 (3.5 bpw) | GGUF | 12.35 | download | — | — | — | — | AIME 2026 96.67 (29/30); GPQA-Diamond 90.4 (179/198), 11 items still pending; MMLU-Pro 87.2 (436/500), 20 items still pending; LiveCodeBench v6 88.6 (155/175), running (phase 5 of 6, 19/22 done); IFBench 78.0 (117/150), 6 items still pending | vs BF16 | L62 @ main |
| Launch80/PARO-int5 [L79-85] | this checkpoint (PARO-int5) | other | 21.00 | download | — | 95.21% | — | — | kld:top256 0.0100; GSM8K 97.40% | vs BF16 | L78 @ main |
| LeaderboardModel1/AutoRound-MXFP4-ModelFree absolute accuracies, no reference comparison [L28-31] | Qwen3.8-27B-AutoRound-MXFP4-ModelFree | MXFP4 | — | — | — | — | — | hellaswag 0.6166; piqa 0.7960 | reference not stated | L30 @ main | |
| LeaderboardModel1/AutoRound-W4A16 No BF16 comparison score on card [L26-31] | AutoRound-W4A16 | AutoRound | — | — | — | — | — | gsm8k 0.9704; hellaswag 0.8252; mmlu 0.8208; piqa 0.8074 | reference not stated | L28 @ main | |
| LeaderboardModel1/AutoRound-W4A16-tuning Scores printed as fractions [L28-31]; no reference model scores on card | Qwen3.8-27B-AutoRound-W4A16-tuning | AutoRound | — | — | — | — | — | gsm8k 0.9704; hellaswag 0.8246; mmlu 0.8253; piqa 0.8112 | reference not stated | L28 @ main | |
| LeaderboardModel1/AutoRound-W4A16-v2 Scores printed as fractions; no baseline comparison [L26-31] | AutoRound-W4A16-v2 | AutoRound | — | — | — | — | — | gsm8k 0.9697; hellaswag 0.8241; mmlu 0.8269; piqa 0.8107 | reference not stated | L28 @ main | |
| leonsarmiento/3bit-mlx 50-question samples, 2pp = 1 question [L88-89]size is Footprint row [L23]3.716 bits per weight stated in prose [L52] | Qwen3.8-27B-3bit-mlx (clean 3-bit) | MLX | 12.71 | download | — | — | — | — | MMLU 76.0%; MMLU_PRO 58.0%; bpw 3.716 bits per weight | other or no reference | L23 @ main |
| leonsarmiento/3bit-mtp-mlx 3.716 bpw base + 0.33 GB MTP stated in prose [L51]Gemma-4 and Qwen3.6 rows are other models, omitted [L80-84] | Qwen3.8-27B 3bit-mtp | MLX | 13.04 | download | — | — | — | — | MMLU 80.0%; MMLU_PRO 61.0%; HUMANEVAL 94.0%; MBPP 70.0% | reference not stated | L26 @ main |
| libertyhhn/GPTQ-W4A16-g128 Weights file 18.6 GB from prose [L24]Paired differences with 95% intervals reported in table [L53-55], not mapped (no matching key) | GPTQ W4A16 g128 (This model) | GPTQ | 18.60 | download | — | — | — | — | ppl:wikitext2 6.28; MMLU 0-shot 84.08; MMLU 5-shot 82.31; GSM8K 96.21 | vs BF16 | L24 @ main |
| Litwein/oQ4e-DWQ-MTP-Vision-MLX Held-out top-K teacher KL vs oQ8e teacher, not task accuracy and not comparable to other datasets/teachers [L99-101]20% reduction vs M4Q init [L130] | oQ4e -> M4Q -> DWQ (this repo) | MLX | — | — | — | — | — | kld:heldout_teacher 0.0229 | other or no reference | L129 @ main | |
| lmcoleman/MagicQuant-GGUF Cell holds ppl and percent increase vs BF16 together, same id used for both [L49]Same quant listed as 20.89 GiB in tier table [L61] | Qwen3.8-27B-Q6_K.gguf | GGUF | 22.40 | download | — | — | — | 6.7579 (+0.20%) | ppl:wikitext-2 6.7579 (+0.20%) | vs BF16 | L49 @ main |
| lmcoleman/MagicQuant-GGUF Cell holds ppl and percent increase vs BF16 together, same id used for both [L48]Same quant listed as 14.65 GiB in tier table [L59] | Qwen3.8-27B-Q4_K_M.gguf | GGUF | 15.70 | download | — | — | — | 6.7611 (+0.25%) | ppl:wikitext-2 6.7611 (+0.25%) | vs BF16 | L48 @ main |
| lmcoleman/ROCmFPX-GGUF Headline note reports MQ-Q6 PPL 6.7412 vs table value [L31,L97] | Qwen3.8-27B-ROCMFPX-MQ-Q6.gguf | GGUF | 22.20 | download | — | — | 6.7579 (+0.20%) ppl:wikitext | — | ppl:wikitext_headline 6.7412 | vs BF16 | L97 @ main |
| lmcoleman/ROCmFPX-GGUF File table cell holds ppl with percent delta [L96]Headline note at L30-32 reports a different, later MQ-Q4 PPL 6.9240 (+2.7%) than table; conflict, see prose [L32] | Qwen3.8-27B-ROCMFPX-MQ-Q4.gguf | GGUF | 15.70 | download | — | — | 6.7611 (+0.25%) ppl:wikitext | — | ppl:wikitext_headline 6.9240 | vs BF16 | L96 @ main |
| Lottolabs/TT-Mixed-BFP4-BFP8-P150 Reference is the TT all-BFP8 build, not BF16 [L48]~22.1 GB checkpoint stated in prose [L110] | GPTQ BFP4 (this checkpoint) vs TT all-BFP8 | NVFP4 | — | — | — | — | — | kld:chat 0.070; top1:chat 0.897; kld:code 0.047; top1:code 0.943; kld:instruct 0.011; top1:instruct 0.967; kld:all 0.055; top1:all 0.925 | vs BF16 | L52 @ main | |
| Lottolabs/TT-Mixed-BFP4-BFP8-P150 KV-cache quantization effect on same weights, not weight quantization; reference is BFP8 KV [L64]Delta NLL +/- SE: book 32K-128K c15, book 262K c18, code c21, long chat c24; top-1 values are fractions | 262K profile 4-bit V cache vs BFP8 KV cache | NVFP4 | — | — | — | — | — | kld:book_32K-128K 0.0074; top1:book_32K-128K 0.957; kld:book_262K 0.0093; top1:book_262K 0.937; kld:code_32K-262K 0.0045; top1:code_32K-262K 0.981; kld:long_chat_61K 0.0062; top1:long_chat_61K 0.966 | vs BF16 | L68 @ main | |
| lribeiro/FP8-Pessoa Body-only KLD against a shared BF16 LM head [L538]Size stated only as ~28 GB in prose/row label [L276]Throughput and composite score also reported [L249] | Pessoa (FP8 W8A16 weight-only, GPTQ) | FP8 | — | 0.005559 | 97.47% | — | — | kld_99 0.085659; kld:multilingual 0.00192; kld:scientific 0.00193; kld:encyclopedic 0.00377; kld:code 0.00596; kld:literary 0.00985 | other or no reference | L187 @ main | |
| Lucebox/IQ4_XS-fast-GGUF requantized from bartowski imatrix IQ4_XS with --pure [L17] | pure IQ4_XS (this file) | GGUF | 14.54 | download | 0.028 | 92.0% | — | +1.9% | — | vs a Q8_0 | L23 @ main |
| lued/INT8-W8A16-DFlash2 [L70-72] | INT8 W8A16 (this build) | INT8 | 30.06 | download | 0.0007 | 98.1% | — | — | — | vs BF16 | L70 @ main |
| lued/INT8-W8A16-MTP size is weight files [L91]long-context probe values from L210 | This W8 / BF16 A16 (INT8 symmetric, groups of 128) | INT8 | 31.61 | download | 0.000894 | 99.36% | — | — | — | vs BF16 | L91 @ main |
| lued/INT8-W8A16-MTP same INT8 weights with FP16 activations [L183] | This checkpoint, FP16 profile (W8 / FP16 A16) | INT8 | — | 0.000641 | 99.19% | — | — | — | vs BF16 | L183 @ main | |
| machinez/d3-NVFP4 scores are for the d3 decision model (a different model from base Qwen3.8-27B), not NVFP4-vs-bf16 quant comparison; card page is d3's, NVFP4 file not distinguished [L151]dagger benchmarks are approximate rebuilds [L171]Perception Test 77.4 internal evaluation in prose [L35]competitor models (Perplexity Decider, Fastino GLiDE, Jev, Torchcast, Decision 2.0, JEV-27B-VL, Solomon) from live board data are different models and omitted [L199] | d3 | NVFP4 | — | — | — | — | — | Jev Decision Index 64.7; Jev Public 65.5; Jev Same-skill tests 62.8; Jev New-domain tasks 56.6; CV-Bench 76.2; BLINK 59.9; RealWorldQA 73.4; CharXiv 91.7; InfographicVQA 96.7; Mind2Web 90.0; Winoground 85.8; R-Bench-M 33.0; MMMU-Pro vision 47.0; Perception Test 77.4 | reference not stated | L151 @ main | |
| magiccodingman/MagicQuant-GGUF Alt imatrix A/B values in L138 table (generic 0.000703 vs Unsloth 0.000705) [L138] | MQ-Q6_K_1 | GGUF | 29.03 | download | 0.000703 | — | — | — | — | reference not stated | L36 @ main |
| magiccodingman/MagicQuant-GGUF Primary KLD is the Unsloth-imatrix rebuild; generic is c81 [L139] | MQ-Q6_K_2 | GGUF | 27.26 | download | 0.000873 | — | — | — | — | reference not stated | L37 @ main |
| magiccodingman/MagicQuant-GGUF Unsloth-imatrix rebuild; generic is c87 [L140] | MQ-Q6_K_3 | GGUF | 25.94 | download | 0.001047 | — | — | — | — | reference not stated | L38 @ main |
| magiccodingman/MagicQuant-GGUF | MQ-Q6_K_4 | GGUF | 23.21 | download | 0.001518 | — | — | — | — | reference not stated | L41 @ main |
| magiccodingman/MagicQuant-GGUF | MQ-Q5_K_1 | GGUF | 22.00 | download | 0.002427 | — | — | — | — | reference not stated | L42 @ main |
| magiccodingman/MagicQuant-GGUF Unsloth-imatrix rebuild [L145] | MQ-Q5_K_2 | GGUF | 20.91 | download | 0.003146 | — | — | — | — | reference not stated | L43 @ main |
| magiccodingman/MagicQuant-GGUF Unsloth-imatrix rebuild [L146] | MQ-Q5_K_3 | GGUF | 20.10 | download | 0.003562 | — | — | — | — | reference not stated | L44 @ main |
| magiccodingman/MagicQuant-GGUF | MQ-Q5_K_S_1 | GGUF | 18.98 | download | 0.004818 | — | — | — | — | reference not stated | L45 @ main |
| magiccodingman/MagicQuant-GGUF | MQ-Q4_K_M_1 | GGUF | 17.62 | download | 0.007412 | — | — | — | — | reference not stated | L46 @ main |
| magiccodingman/MagicQuant-GGUF Unsloth-imatrix rebuild [L150] | MQ-IQ4_XS_1 | GGUF | 16.34 | download | 0.013723 | — | — | — | — | reference not stated | L48 @ main |
| magiccodingman/MagicQuant-GGUF | MQ-IQ2_M_1 | GGUF | 11.96 | download | 0.057811 | — | — | — | — | reference not stated | L53 @ main |
| magiccodingman/MagicQuant-GGUF | MQ-IQ2_M_2 | GGUF | 10.77 | download | 0.092394 | — | — | — | — | reference not stated | L55 @ main |
| magiccodingman/MagicQuant-GGUF Unsloth-imatrix rebuild [L160] | MQ-IQ2_XXS_1 | GGUF | 8.3 | download | 0.270304 | — | — | — | — | reference not stated | L58 @ main |
| maglun/MLX-Mixed-3.80bpw PPL values from the two tables use different corpora and are not interchangeable [L91-93]n=8 and n=12 samples, screening-tier only [L119-122]decode tok/s is not a quality metric [L101] | mixed affine 3.80 bpw (this release) | MLX | 13.70 | download | — | — | — | — | ppl:heldout-4doc 6.4916; ppl:heldout-6doc 6.2212; HumanEval+ 75.0%; GSM8K 91.7% | vs BF16 | L74 @ main |
| maglun/MLX-Mixed-4.95bpw re-measured on Apple M5 Pro; launch-day M4 Pro PPL was 5.9028 [L102-103]tiny n; HumanEval+ does not discriminate [L115]size is full tensor payload incl. BF16 vision [L37] | this release - mixed affine 4.95 bpw | MLX | 17.57 | download | — | — | 5.9056 | — | HumanEval+ 75.0%; GSM8K 100.0% | vs BF16 | L94 @ main |
| mahadev9/fp8 mean_kld and top1_pct are the 'overall' bucket [L29] | fp8 | FP8 | — | 0.2244 | 91.92% | — | — | kld:prose 0.0043; top1:prose 96.78%; kld:agentic 0.2864; top1:agentic 90.55% | vs BF16 | L27 @ main | |
| majentik/MLX-2bit [L43,L44] | MLX 2bit (affine, group size 32) | MLX | — | — | — | — | — | arc_easy_acc 0.4900; hellaswag_acc 0.4100 | reference not stated | L43 @ main | |
| majentik/MLX-3bit No full-precision baseline shown for these scores [L41-44] | Qwen3.8-27B-MLX-3bit | MLX | — | — | — | — | — | arc_easy_acc 0.5850; hellaswag_acc 0.4850 | reference not stated | L43 @ main | |
| majentik/MLX-4bit no BF16 comparison or sample size given [L39-45] | Qwen3.8-27B-MLX-4bit (affine, group size 32) | MLX | — | — | — | — | — | arc_easy_acc 0.5900; hellaswag_acc 0.5150 | other or no reference | L43 @ main | |
| majentik/MLX-5bit no sample size or harness stated [L40-44] | MLX 5-bit (affine, group size 64) | MLX | — | — | — | — | — | arc_easy_acc 0.6050; hellaswag_acc 0.5350 | other or no reference | L43 @ main | |
| majentik/MLX-6bit scores are round to two decimals (0.6000, 0.5400), sample size not stated [L43-44]no BF16 baseline given on card | Qwen3.8-27B-MLX-6bit (affine, group size 64) | MLX | — | — | — | — | — | arc_easy_acc 0.6000; hellaswag_acc 0.5400 | reference not stated | L43 @ main | |
| majentik/MLX-8bit no base-model comparison; small-sample scores [L41-L44] | MLX 8bit | MLX | — | — | — | — | — | arc_easy_acc 0.6150; hellaswag_acc 0.5400 | reference not stated | L43 @ main | |
| majentik/MLX-MXFP4 mxfp4, 4-bit, group size 32 [L14,L23]no reference scores given [L43-44] | MLX MXFP4 | MLX | — | — | — | — | — | arc_easy_acc 0.6300; hellaswag_acc 0.5000 | reference not stated | L43 @ main | |
| malaiwah/EXL3-EDA-research worse than hydrated by +0.000366 mean KLD, 95% CI [+0.000334, +0.000398] [L28]serialized size 21,586,964,548 B identical to hydrated [L30]; bytes cell c29 belongs to another tablemedian/p99.9/exact max columns not mapped (p99.9 is not p99) [L95] | error-driven allocation (this build) | EXL3 | — | 0.003066179635178366 | 97.506 % | — | — | — | other or no reference | L95 @ main | |
| malaiwah/EXL3-K5K6 resident 20.32 GiB (c9) [L194]download is BF16 attention encoded at load [L482-484]v5 body-only KL, not served-logit or task-quality [L36-39]max single position 22.241 worst of five [L288] | K5K6 (this build, v5 10M) | EXL3 | 30.60 | download | 0.003210 | 97.52 % | — | — | — | vs BF16 | L194 @ main |
| malaiwah/EXL3-K5K6 shard-0 [L456] | K5K6 this build (shard-0 replay) | EXL3 | — | 0.003141 | 97.61 % | — | — | — | vs BF16 | L456 @ main | |
| malaiwah/EXL3-K5K6-context v3 development suite 136 contexts [L379-390] | this build, BF16 input, MTP off (v3) | EXL3 | 20.73 | resident | 0.009673 | 96.81 % | — | — | — | vs BF16 | L392 @ main |
| malaiwah/EXL3-K5K6-context v5 value is the serialized checkpoint, no int8 input overlay, body-only [L229,L41]size is the download size from the overview table [L139], not from the fidelity tableshard-0 p99.9 c92 (0.1632) is p99.9 not p99, not mapped [L466]MMLU-Pro 58/70 [L742] | this build (context edition), v5 10M | EXL3 | 20.70 | download | 0.003509 | 97.44 % | — | — | kld:v3_corrected 0.009459; kld:v4 0.003990; top1:v4 97.36 %; MMLU-Pro 58/70 (82.9 %); kld:shard0 0.003409; top1:shard0 97.55 % | vs BF16 | L139 @ main |
| malaiwah/EXL3-K5K6-context v3 development suite [L379-390] | this build, int8 input overlay, MTP off (v3) | EXL3 | 19.47 | resident | 0.009738 | 96.80 % | — | — | — | vs BF16 | L393 @ main |
| malaiwah/EXL3-K5K6-hydrated v5 10M body-only, headline [L24, L320]download 21.61 GB [L145] | this build (hydrated) v5 10M | EXL3 | 21.81 | resident | 0.002760 | 97.70 % | — | — | — | other or no reference | L320 @ main |
| malaiwah/EXL3-K5K6-hydrated prior receipt, different suite, not comparable to v5 [L446-448, L777] | this build (hydrated) v3 136-context | EXL3 | 21.81 | resident | 0.007406 | 97.19 % | — | — | — | other or no reference | L779 @ main |
| malaiwah/EXL3-K5K6-hydrated shard 0 only, vLLM [L495]p99.9 0.1313 not mapped (no key) | this build (hydrated) v5 shard0 cross-engine table | EXL3 | 21.60 | download | 0.002700 | 97.80 % | — | — | — | other or no reference | L495 @ main |
| malaiwah/EXL3-K5K6-hydrated [L853] | this build (hydrated) v4 36-context | EXL3 | — | 0.003093 | 97.63 % | — | — | — | other or no reference | L853 @ main | |
| malaiwah/EXL3-K6-parity p99.9 (c11) and max (c12) not mapped: no matching key [L109] | this build (K6 parity, EXL3) | EXL3 | 23.03 | download | 0.001634 | 98.248 % | — | — | — | vs BF16 | L109 @ main |
| malaiwah/EXL3-S16-V-research 12.7922 GiB whole tree as built [L89]rejected: 1.5x the pre-registered 0.030 threshold [L26] | S16-V (this build) | EXL3 | 13.74 | download | 0.045374 | 91.73 % | — | — | kld_99 0.543578 | vs BF16 | L89 @ main |
| malekoo/MLX-4bit +2.2% relative to bf16 [L33]4.501 effective bits/weight in prose [L22] | MLX 4-bit (group size 64, affine) | MLX | 15.00 | download | — | — | — | — | ppl:wikitext2 7.0871 | vs BF16 | L22 @ main |
| malekoo/MLX-8bit size ~29 GB on disk; 8.501 effective bits/weight in prose only [L22]decode speed and peak memory also reported, not quality [L33] | MLX 8-bit (group size 64, affine) | MLX | 29.00 | download | — | — | — | +0.14% | ppl:wikitext-2 6.9447 | vs BF16 | L22 @ main |
| malekoo/MLX-bf16 lossless bf16 conversion, ~54 GB on disk [L31]peak memory ~54 GB on M5 Max [L126] | Qwen3.8-27B-MLX-bf16 | MLX | 54.00 | download | — | — | 6.9352 | — | — | other or no reference | L31 @ main |
| manitcor/Obliterated-E03-bnb-4bit KLD is first-token only over 33 prompts, max 0.3652 [L144-145]Refusal/coherence rows describe abliteration behavior [L138-139]Peak CUDA 22.74 GB, steady 17.66 GB on A100 reload [L153,L197] | NF4 artifact | bitsandbytes | — | — | — | 3.4106 | 6.25% | capability checks 83.3% | vs BF16 | L140 @ main | |
| mauromedda/fucina-qwen38-v2 Decode throughput table is speed only [L105-110]Not promoted for general production or multimodal quality [L122] | custom NVFP4 local-Hessian v2 packed | other | — | — | — | — | — | tool-eval-bench 86/100 | reference not stated | L115 @ main | |
| MESHIVEAI/Qwen-3.8-27B-Q4_K_M-Imatrix wikitext PPL and KLD on Wikitext-2 test [L50-L58]delta PPL vs F16 is absolute, not percent [L59]size is weights-only from prose [L126]imatrix calibrated on Wikitext-2 train, same domain as eval [L233-L234] | Q4_K_M imatrix (this model) | other | 16.80 | text weights (excl. vision, MTP) | 0.017486 ± 0.000201 | 94.18% ± 0.06 | 6.9619 | — | ppl:korean-wikipedia 6.9485; ppl_delta_pct:korean-wikipedia +1.02% | other or no reference | L54 @ main |
| Mia-AiLab/EXL3-3.5bpw tool-eval-bench 87-88/100 given as range in prose [L149]needle test: 10%/50% depth pass at 300k [L107-109] | Qwen3.8-27B-EXL3-3.5bpw | EXL3 | 14.20 | unknown | — | — | — | — | tool-eval-bench 87–88 / 100 | other or no reference | L19 @ main |
| MiawTeam/Uncensored-GGUF only file with a PPL value in Files table [L77-89] | Qwen3.8-27B-Uncensored-Q4_K_M.gguf | GGUF | 16.80 | download | — | — | — | — | ppl:wikitext-2 PPL = 7.1814 +/- 0.25227 | other or no reference | L80 @ main |
| MiawTeam/Uncensored-GGUF benchmarks are on bf16 abliterated weights, not on the GGUF quants [L152-153]KL is first-token divergence from the base caused by abliteration, not quantization error [L208]0-shot, deltas within standard error [L152-155] | Qwen3.8-27B-Uncensored (bf16, abliterated) | GGUF | — | — | — | — | — | MMLU 83.3; ARC-Challenge 57.7; HellaSwag 82.9; Winogrande 75.3; kld:first-token-vs-base 0.1191 | other or no reference | L146 @ main | |
| Mihai-LeanZero/Atlassian-Q6_K-GGUF Same top-1 token 98.5 [L33]; maximum KLD 6.626617 not mappedModel is an Atlassian-adapter fine-tune; KLD is Q6_K vs its own bf16 GGUF [L33] | Q6_K | GGUF | — | 0.00306 | 98.5 | — | — | — | vs BF16 | L33 @ main | |
| Mihai-LeanZero/Atlassian-Q8_0-GGUF maximum KLD 10.87871 also stated [L33]app-compile counts from llama-server are task scores, not mapped [L37] | Q8_0 | GGUF | — | 0.000842 | 99.374 | — | — | — | vs BF16 | L33 @ main | |
| Minachist/INT5-Flat-5.7bpw-AutoRound Stated as 5.73 bpw [L18]mean KL cell includes SE [L31] | INT5-Flat 5.7bpw | AutoRound | — | — | — | 7.9638 ppl:wikitext | — | kld:wikitext 0.008392 (±SE 0.000113); top1:wikitext 95.980 %; kld_99 0.079737 | vs BF16 | L31 @ main | |
| Minachist/INT6-Flat-6.6bpw-AutoRound 6.60 bpw stated in prose [L18]; card warns KLD not comparable across methods [L26]p99.9 KL 0.095758 [c11], max 2.923414 [c12] | INT6-Flat 6.6bpw | AutoRound | — | 0.003051 (±SE 0.000047) | 97.471 % | 7.9294 | — | kld_99 0.025704 | vs BF16 | L32 @ main | |
| Minachist/INT6-Mixed-AutoRound bpw 7.30 stated in prose [L20]loaded weights 12.19 GiB per GPU under TP2, not a total size [L123]median/p90/p95/p99.9/max KL also reported [L36-41] | INT6-Mixed | AutoRound | — | 0.001980 (±SE 0.000051) | 97.905 % | 7.9183 | — | kld_99 0.014824 | vs BF16 | L35 @ main | |
| Minachist/INT8-AutoRound ln(PPL ratio) given [L85] | linear-attn-bf16-gs128 | AutoRound | 36.83 | download | — | — | — | — | kld:wikitext 0.000956 ± 0.000021; top1:wikitext 98.59%; kld_99 0.00610 | vs BF16 | L28 @ main |
| Minachist/INT8-AutoRound ln(PPL ratio) given [L85] | linear-attn-bf16 | AutoRound | 36.40 | download | — | — | — | — | kld:wikitext 0.00110 ± 0.00001; top1:wikitext 98.47%; kld_99 0.00719 | vs BF16 | L26 @ main |
| Minachist/INT8-AutoRound ln(PPL ratio) given [L85] | main-gs128 | AutoRound | 31.46 | download | — | — | — | — | kld:wikitext 0.00101 ± 0.00002; top1:wikitext 98.52%; kld_99 0.00666 | vs BF16 | L27 @ main |
| Minachist/INT8-AutoRound ln(PPL ratio) given, not a plain ratio [L85]PPL base 7.9112 to quantized 7.9101 [L84] | main | AutoRound | 30.92 | download | — | — | — | — | kld:wikitext 0.00124 ± 0.00002; top1:wikitext 98.40%; kld_99 0.00825 | vs BF16 | L25 @ main |
| mlx-bun/Trellis-3.2bpw Size 11.99 GB is from prose [L67]; complete download with MTP companion 12.25 GB [L69]Card states 3.23 bpw [L65]Differences vs BF16 compare complete inference configurations, not weight quantization alone [L111]Thirteen GPQA responses hit the token limit and scored incorrect [L85] | Qwen3.8-27B-Trellis-3.2bpw | MLX | 11.99 | download | — | — | — | — | GSM8K-Platinum_flexible 96.28%; GSM8K-Platinum_strict 96.11%; GPQA Diamond 72.73% | vs BF16 | L67 @ main |
| mlx-community/OptiQ-4bit table says 19.0 GB, later prose says 18.9 GB on disk [L39,L85]mixed precision: 261 layers 8-bit, 237 layers 4-bit [L32-L33] | OptiQ-4bit | MLX | 19.00 | download | — | — | — | — | MMLU 84.1%; GSM8K 96.1%; IFEval 77.1%; BFCL-V3-simple 93.5%; HumanEval 92.1%; HashHop 85.0%; Capability-Score 87.98 | reference not stated | L39 @ main |
| mtrskeen/ninfer-minq4-mtpq4-visionq4-5080 PPL +0.04 is simulated and relative to the IQ3_XXS anchor, not BF16 [L50]RULER NIAH retrieval 3/3 at each of 8K/32K/64K [L94]Speed and VRAM figures omitted | qwen3_8_27b_minq4_mtpq4_visionq4.ninfer | other | — | — | — | — | — | gsm8k_cot 0.96 | other or no reference | L50 @ main | |
| mudler/APEX-GGUF comparison uses trunk-only files; published Mini is 13.94 GB (c1), about 0.45 GB larger with MTP head [L32,L69] | APEX-I-Mini (trunk-only comparison build) | GGUF | 13.49 | download | 0.049020 | — | — | — | — | vs BF16 | L62 @ main |
| mudler/APEX-GGUF published Nano is 11.24 GB (c3) [L33,L69] | APEX-I-Nano (trunk-only comparison build) | GGUF | 10.79 | download | 0.121640 | — | — | — | — | vs BF16 | L63 @ main |
| MuXodious/absolute-heresy.rc12-GGUF piqa 0.4951 and winogrande 0.4957 are near chance level, as printed [L16,L23]refusals 2/101 [L13]chosen release candidate [L38] | RC1 (Trial 377) Q4_K_M | GGUF | — | — | — | — | — | piqa_acc 0.4951; piqa_acc_norm 0.4951; copa 0.55; winogrande 0.4957 | other or no reference | L13 @ main | |
| MuXodious/absolute-heresy.rc12-GGUF refusals 6/101 [L25]benchmark values identical to RC1 [L28-35] | RC2 (Trial 318) Q4_K_M | GGUF | — | — | — | — | — | piqa_acc 0.4951; piqa_acc_norm 0.4951; copa 0.55; winogrande 0.4957 | other or no reference | L25 @ main | |
| nerkyor/Uncensored-Heretic-T10-BF16 Scores are for the Q8_0 GGUF of an uncensored/Heretic-abliterated model, not compared against base Qwen3.8-27B [L52-54]Refusal140 hard refusal 1/140 c4, soft disclaimer 32/140 c5 [L59-60]LCB difficulty easy 23/23, medium 30/31, hard 23/46 [L64] | Frozen Q8_0 results | other | — | — | — | — | — | MMLU500 435/500 (87.0%); GPQA198 169/198 (85.35%) | other or no reference | L56 @ main | |
| neroued/nvfp4-NInfer size is resident MTP3 weights 19.729 GiB [L259]IFBench is prompt-level strict [L310]; one sample per problem | NInfer NVFP4 | NVFP4 | 21.18 | resident | — | — | — | — | GPQA-Diamond 90.40%; ERQA 66.25%; RealWorldQA 83.53% | vs BF16 | L259 @ main |
| nibor1896/CNQ4.5 median KLD (c56) not mapped, no key [L315]kld_99 is the p99 of per-token KLD [L315] | Qwen3.8-27B-CNQ4.5 (container, f32 math) | other | 17.84 | download | 0.2233 ± 0.0190 | 86.05 % | — | — | kld_99 2.92; kld:code 0.4435; kld:prose 0.1456 | vs BF16 | L32 @ main |
| nicosuter/AWQ weights size before KV cache [L194]; recovery has 95% CI; scores pooled over two passes for GPQA, LCB, MMMU-Pro [L105-107]; table does not report delta ids here (deltas c22,c27,c32,c37,c42,c46,c51) | Qwen3.8-27B-AWQ | AWQ | 21.50 | download | — | — | — | — | BFCL 81.33; GPQA_Diamond 88.64; LiveCodeBench_v6 89.14; MMMU-Pro 75.84; Multimodal 86.34; macro_5_suites 84.26; RULER 92.90 | other or no reference | L112 @ main |
| nm-testing/MXFP6A8 calibration-free compressed-tensors quant [L16] | Qwen3.8-27B-MXFP6A8 | other | — | — | — | — | — | GSM8K 0.770 | vs BF16 | L39 @ main | |
| notnaurw/oQ5e-SAGE-ARCS-mtp kld:agentic and kld:stock are truncated head-KL (top-1024), not full-vocab KL [L404,L409,L506]agentic head-KL 0.3166 is a pooled figure, corpora not comparable with each other [L261,L433]size 19.68 GiB (21.13 GB) printed in headline, basis not stated [L79]ppl ratios 0.993 and 0.997 cover both builds, order not stated, omitted [L336] | SAGE-ARCS | other | 21.13 | unknown | — | — | — | — | kld:agentic 0.3166; kld:stock 0.014518; kld:agentic_fullvocab 0.352813 | vs BF16 | L79 @ main |
| notnaurw/oQ5e-SAGE-CREST-mtp Main tables are not in the candidate list; only prose values captured [L197,L390,L644,L652]Agentic head-KL is corpus-dependent and not comparable to generalist head-KL [L577] | SAGE-CREST | other | 22.50 | download | — | — | — | — | kld:generalist_head 0.010114; kld:agentic_full_vocab 0.268643; kld:agentic_head 0.2433 | vs BF16 | L80 @ main |
| notnaurw/oQ5e-SAGE-PATHS-mtp head-KL values are not comparable across corpora [L573]agentic-basis head-KL 0.2794 and full-vocab KL 0.310274 given in prose only [L396,L637]vision tower removed; text-only [L72]bpw counted over whole checkpoint incl. bf16 MTP head [L104] | SAGE-PATHS | other | 20.96 | download | — | — | — | — | kld:generalist_head 0.012280; kld:agentic_head 0.2794; kld:agentic_full_vocab 0.310274 | vs BF16 | L84 @ main |
| notnaurw/oQ5e-SAGE-mtp mean_kld is head-KL over bf16 top-1024 logits, not full-vocab KL [L200]top-1 agreement is a fraction (0.9523), not a percent [L264]per-rung HTML tables (stock, ladder, top-5, ppl ratio) have no candidate ids; only prose values mapped21.13 GB (19.68 GiB) shipped size incl. bf16 MTP head [L77] | SAGE oQ5e mtp (shipped build) | other | 21.13 | download | 0.010348 | 0.9523 | — | — | kld:full_vocab 0.011516 | vs BF16 | L77 @ main |
| nova-quill/OBLITERATUS-OBLITERATED-unsloth-mlx MMLU is for the abliterated weights; card does not say which format (bf16 vs MLX/GGUF quant) was evaluated [L112-116]stderr +/-0.65, n=2,850 [L116] | V2 (s78, blend) | MLX | — | — | — | — | — | MMLU 84.32% | other or no reference | L116 @ main | |
| okayuji/coreml-agent agent benchmark 48/48 runs passed [L124]top-1 agreement vs bf16 633/640 (98.9%), 0 hard flips [L256-258]int4 build (unpublished): 11 hard top-1 flips in 640 positions vs 0 for int8 [L238-240] | int8 Core ML bundle | other | 30.30 | download | — | 633/640 (98.9%) | — | — | — | vs BF16 | L124 @ main |
| Omega-Factory/GGUF 1 seed [L60] | Q8_0 | GGUF | 29.05 | download | — | — | — | — | ARC-MC 94.5% (Δ 0.0 pp); GSM8K 63.5% (Δ +4.5 pp); MATH-500 22.0% (Δ -1.0 pp) | other or no reference | L34 @ main |
| Omega-Factory/GGUF GSM8K drop -25.5 pp vs FP16 [L35]1 seed [L61] | Q4_K_M | GGUF | 16.81 | download | — | — | — | — | ARC-MC 95.0% (Δ +0.5 pp); GSM8K 33.5% (Δ -25.5 pp); MATH-500 17.5% (Δ -5.5 pp) | other or no reference | L35 @ main |
| Omega-Factory/GGUF GSM8K 86.0% exceeds FP16 59.0% and Q8_0, as printed [L36]1 seed [L62] | Q2_K | GGUF | 10.86 | download | — | — | — | — | ARC-MC 93.0% (Δ -1.5 pp); GSM8K 86.0% (Δ +27.0 pp); MATH-500 10.0% (Δ -13.0 pp) | other or no reference | L36 @ main |
| oracomputing/OQ3-GGUF retention 95.4% (c15) [L47]size from Files table [L35]; body BPW c14 | OQ3 (this repo) | GGUF | 10.54 | download | — | — | — | — | MMLU-Pro 73.94; AIME-25 90.42; Average 82.18 | other or no reference | L35 @ main |
| orcarouter/OrcaSAQ-2-27B WikiText-2 table [L116-119]3.21 bpw is decoder average [L273] | OrcaSAQ2 27B | other | 12.30 | download | 0.031 | 93.2% | 5.6482 | +0.02% | — | vs BF16 | L97 @ main |
| orcarouter/MLX KLD p95 0.00233 in table [L112] not mapped (no key)group size 32 | 8-bit | MLX | 31.00 | download | 0.00068 | 98.34 % | — | — | — | vs BF16 | L95 @ main |
| orcarouter/MLX KLD p95 in table [L113] not mappedgroup size 32 | 6-bit | MLX | 24.00 | download | 0.00216 | 96.97 % | — | — | — | vs BF16 | L96 @ main |
| orcarouter/MLX KLD p95 in table [L114] not mappedgroup size 64 | 4-bit | MLX | 16.00 | download | 0.02824 | 92.68 % | — | — | — | vs BF16 | L97 @ main |
| OS-Software/Hemmingway-1-Uncensored-Heretic-GGUF KL divergence is of the abliterated model vs the original, not a quantization metric [L45]refusal counts 0/100 vs 95/100 omitted [L44]no quant type or size stated in candidates | This model (Heretic decensored) | GGUF | — | 0.0474 | — | — | — | — | vs BF16 | L45 @ main | |
| OS-Software/MTP-heretic-ja-GGUF KL divergence measures abliteration drift from unsloth/Qwen3.8-27B, not quantization loss [L33]refusal keywords 0/100 vs original 100/100 [L32], not mapped | decensored Qwen3.8-27B (heretic-ja) | GGUF | — | 0.0528 | — | — | — | — | vs BF16 | L33 @ main | |
| OS-Software/Uncensored-Heretic-v2-GGUF KL divergence here measures the abliterated model vs the original, not quantization loss [L33]Refusal keywords score is 0/100 for this model vs 98/100 original; not a fidelity metric [L32]No per-quant GGUF sizes or quant levels listed in the card | Uncensored-Heretic-v2 (this model) | GGUF | — | 0.0182 | — | — | — | — | other or no reference | L33 @ main | |
| OS-Software/Uncensored-Heretic-v3-GGUF KL divergence vs original model as reported by the abliteration tool [L40]Refusals 0/100 vs 99/100 for original (not a quality metric key) [L39]Upstream Qwen3.8-27B benchmark tables pasted from original card are not measurements of this model [L172-499] | This model (Heretic abliterated) | GGUF | — | 0.0087 | — | — | — | — | vs BF16 | L40 @ main | |
| oshiteku/Sanshoku-GGUF Small-scale local evaluation [L66]; top-1 and KL averaged over chat and coding-agent prompts, thinking and non-thinking [L76]Per-file sizes: Q2_0 7.70, PQ2_0 7.28, PTQ1_0 6.00 GB [L35-37]; eval table does not say which file | Sanshoku | GGUF | 7.3 | download | 0.297 | 84.4% | — | — | SWE-bench Verified Mini 18% | vs BF16 | L70 @ main |
| pearsonkyle/GPTQ-W4A16 KLD values are medians, range 0.0021-0.0205 across 6 distributions; lowest on agentic, highest on general English [L104,L109]5.2 bpw overall, trunk 4.05 bpw [L60,L127]vision 3/3 probe [L106] | W4A16 (GPTQ int4 g128 sym) | GPTQ | 19.43 | download | — | — | — | — | SchemaValidRate 0.948; MalformedToolCalls 0; kld:agentic 0.0021; kld:general_english 0.0205 | vs BF16 | L42 @ main |
| pearsonkyle/imatrix-MTP-GGUF KLD is median [L89]size includes bundled Q8_0 MTP head [L80]agentic rows are a band within +-3.8pp, not a ranking [L108] | Q5_K_M | GGUF | 19.68 | download | — | — | — | — | ppl:external 23.274; kld:external 0.0036; top1:external 91.8%; kld:general 0.0062; kld:tools 0.0014; kld:agentic 0.0006; kld:broad 0.0076; kld:cal8k 0.0025; Tool-selection acc 0.474; Param accuracy 0.265 | other or no reference | L97 @ main |
| pearsonkyle/imatrix-MTP-GGUF recommended default [L69]KLD is median [L89] | IQ4_XS | GGUF | 15.54 | download | — | — | — | — | ppl:external 31.118; kld:external 0.0105; top1:external 88.2%; kld:general 0.0154; kld:tools 0.0032; kld:agentic 0.0013; kld:broad 0.0204; kld:cal8k 0.0062; Tool-selection acc 0.503; Param accuracy 0.274 | other or no reference | L97 @ main |
| pearsonkyle/imatrix-MTP-GGUF KLD is median [L89] | IQ3_M | GGUF | 13.04 | download | — | — | — | — | ppl:external 37.363; kld:external 0.0354; top1:external 82.4%; kld:general 0.0569; kld:tools 0.0115; kld:agentic 0.0048; kld:broad 0.0702; kld:cal8k 0.0261; Tool-selection acc 0.454; Param accuracy 0.239 | other or no reference | L97 @ main |
| pearsonkyle/imatrix-MTP-GGUF KLD is median [L89]tool-selection acc cell for IQ2_M not in candidates [L102] | IQ2_M | GGUF | 10.46 | download | — | — | — | — | ppl:external 56.540; kld:external 0.1242; top1:external 72.0%; kld:general 0.2199; kld:tools 0.0498; kld:agentic 0.0169; kld:broad 0.2691; kld:cal8k 0.0840; Param accuracy 0.260 | other or no reference | L97 @ main |
| penkia/TernaryQuench-GGUF download 10.86 GB [L40]second-half PPL scores last 256 tokens of each chunk [L59]1.55% relative RMS weight error from packing [L107] | TernaryQuench 27B (Q2_K) | GGUF | 10.86 | download | — | — | — | — | ppl:wikitext2 12.8569; ppl_second_half:wikitext2 10.0972; PIQA 78.78; ARC-Easy 78.96; ARC-Challenge 55.29; HellaSwag 73.96; WinoGrande 74.43 | vs BF16 | L40 @ main |
| penkia/TernaryQuench-MLX size is download size [L42]zero-shot accuracies in L75 are for the full-ternary training checkpoint, not this MLX artifact [L68-70] | TernaryQuench 27B (MLX) | MLX | 10.53 | download | — | — | — | — | ppl:wikitext2 12.8492 | reference not stated | L42 @ main |
| penkia/TernaryQuench-MLX not the MLX artifact [L68-70] | TernaryQuench 27B (full-ternary training checkpoint, lm-eval) | MLX | — | — | — | — | — | PIQA 78.51; ARC-Easy 79.29; ARC-Challenge 54.01; HellaSwag 73.48; WinoGrande 72.93 | reference not stated | L75 @ main | |
| perkel/MC In-VRAM 23.58 GiB (c18) [L53] | MCXXL (qwen3.8-27b-xxl.mcapy) | other | 28.20 | download | 0.0051 | 98.3% | — | — | kld:reasoning 0.0007; kld:math 0.0008; kld:code 0.0069; kld:chat 0.0151; kld:wiki 0.0021 | vs BF16 | L53 @ main |
| perkel/MC In-VRAM 18.66 GiB (c14) [L52] | MCLarge (qwen3.8-27b-large.mcapy) | other | 22.90 | download | 0.0081 | 97.7% | — | — | kld:reasoning 0.0009; kld:math 0.0010; kld:code 0.0075; kld:chat 0.0283; kld:wiki 0.0026 | vs BF16 | L52 @ main |
| perkel/MC In-VRAM 15.39 GiB (c10) [L51] | MCMedium (qwen3.8-27b-medium.mcapy) | other | 19.50 | download | 0.0298 | 95.2% | — | — | kld:reasoning 0.0058; kld:math 0.0078; kld:code 0.0290; kld:chat 0.0874; kld:wiki 0.0190 | vs BF16 | L51 @ main |
| perkel/MC In-VRAM 13.67 GiB (c6) [L50] | MCSmall (qwen3.8-27b-small.mcapy) | other | 17.60 | download | 0.0478 | 93.8% | — | — | kld:reasoning 0.0107; kld:math 0.0138; kld:code 0.0424; kld:chat 0.1456; kld:wiki 0.0262 | vs BF16 | L50 @ main |
| perkel/MC In-VRAM 12.70 GiB also given (c2) [L49] | MCTiny (qwen3.8-27b-tiny.mcapy) | other | 16.60 | download | 0.0582 | 92.7% | — | — | kld:reasoning 0.0138; kld:math 0.0156; kld:code 0.0471; kld:chat 0.1750; kld:wiki 0.0395 | vs BF16 | L49 @ main |
| Pilcothink/MixedInt2-ultra-lite Custom mixed-precision INT2, group sizes 32/64/128 [L46-47]Recovery rate 93.78% MMLU, 59.42% GSM8K flexible, 50.22% GSM8K strict [L62-63,L85] | Qwen3.8-27B-MixedInt2-ultra-lite (flexible) | other | — | — | — | — | — | MMLU 78.30%; GSM8K flexible 43.29%; GSM8K strict 35.33% | other or no reference | L62 @ main | |
| Pilcothink/MixedInt4-AutoRound ARC-Challenge and MMLU Humanities quantized cells missing from table (difference 0.00 pp, recovery 100%) [L60,L72]recovery rates e.g. MMLU 99.50% (c6), avg 99.38% (c33)group size 32 [L43] | Qwen3.8-27B-MixedInt4-AutoRound | AutoRound | — | — | — | — | — | MMLU 83.07%; BoolQ 80.49%; HellaSwag 82.40%; PIQA 81.66%; WinoGrande 76.16%; Average 76.97% | other or no reference | L58 @ main | |
| PocketAiHub/Abliterated-MTPLX-Optimized-Speed KL measures abliteration drift vs regular MTPLX, not quantization loss vs BF16 [L118]size is download payload 21.31 GB [L68]median, p95, per-suite KLs stated in prose but without ids; speed tables not mapped | Abliterated MTPLX Optimized Speed | other | 21.31 | download | — | 85.94 | — | — | — | vs BF16 | L68 @ main |
| poisonxa/PXQN-OneCard ppl_ratio is assistant-token perplexity vs Q8_0 [L118]KLD is vs Q8_0, not bf16 [L96] | PXQN OneCard (this file) | other | 13.56 | download | 0.0213 | 95.1% | — | — | ppl:wikitext-2 6.979 | vs a Q8_0 | L104 @ main |
| PollardWeights/Penjing-27B-Pollard Recommended default; ik_llama.cpp [L58] | Penjing-27B-IQ2_KT.gguf | other | 7.4 | download | 0.4217 | 79.10% | — | +23 | ppl:wikitext2 5.0817 | other or no reference | L70 @ main |
| PollardWeights/Penjing-27B-Pollard | Penjing-27B-IQ2_XXS.gguf | other | 7.2 | download | 0.5142 | 76.38% | — | +31 | ppl:wikitext2 5.4312 | other or no reference | L69 @ main |
| PollardWeights/Penjing-27B-Pollard Smallest rung; runs in ik_llama.cpp [L60] | Penjing-27B-IQ1_KT.gguf | other | 6.5 | download | 0.5513 | 75.88% | — | +40 | ppl:wikitext2 5.7864 | other or no reference | L68 @ main |
| primitive-ai/mixed-NVFP4-FP8 table header prints 'G'; prose [L45] states GiBsingle-leg row [L90] is the same repo, not a separate build | this repo | NVFP4 | 22.23 | download | — | — | — | — | overall 88.8; knowledge 90.4; call 85.5; abstain 54.5 | vs BF16 | L85 @ main |
| ProCreations/bonsai-2-27b-gsq-rco-gguf PPL stated 5.84% lower than original Bonsai (sign is lower) [L35]Paired accuracy changes vs original Bonsai all include zero in 95% bootstrap interval [L36,L50-55]File 6,496,332,448 bytes (6.496 GB) stated in prose [L21] | Bonsai GSQ + RCO | GGUF | — | — | — | — | — | ppl:wikitext2 8.1627; ARC-C 60.67; ARC-E 80.13; HellaSwag 77.80; Winogrande 74.11 | vs BF16 | L35 @ main | |
| pugant/MTP-Q3_0_ROCMFPX PPL and delta share one cell [L82] | agent Q3_0_ROCMFPX_AGENT | GGUF | 19.54 | download | — | — | — | — | ppl:en 6.7665 (+1.89%); ppl:it 11.9541 (+2.04%); ppl_delta_pct:en 6.7665 (+1.89%); ppl_delta_pct:it 11.9541 (+2.04%) | vs BF16 | L82 @ main |
| pugant/MTP-Q3_0_ROCMFPX PPL and delta share one cell, e.g. '6.9943 (+5.32%)' [L81]effective bpw 4.44 vs nominal 3.50 [L146] | base Q3_0_ROCMFPX | GGUF | 15.17 | download | — | — | — | — | ppl:en 6.9943 (+5.32%); ppl:it 12.3121 (+5.09%); ppl_delta_pct:en 6.9943 (+5.32%); ppl_delta_pct:it 12.3121 (+5.09%) | vs BF16 | L81 @ main |
| quark75/EXL3-3.0bpw ~13 GB on disk [L19] | EXL3 3.0bpw | EXL3 | 13.00 | download | 0.0487 | 91% | 6.985 | — | — | vs BF16 | L19 @ main |
| quark75/Q4_0.gguf ~15 GB file size from prose [L15]~4.5 bits/weight effective stated in prose only [L14]values carry +/- error terms [L24-27] | Qwen3.8-27B-Q4_0 | GGUF | 15.00 | download | 0.0400 +/- 0.0004 | 91.4% +/- 0.07% | 7.0892 +/- 0.0460 | — | — | vs BF16 | L15 @ main |
| QUASAR-QAT/QUASAR-NVFP4 496/496 linears NVFP4 [L82] | QUASAR (this model) | NVFP4 | 19.70 | download | — | — | — | — | GPQA-Diamond 90.91; AIME'26 100.0; MRCR 79.1; MRCR 8-needle 52.7; HMMT'26 73.4; SuperGPQA 62.4; LiveCodeBench 79.2; RULER 32K 96.3; RULER 64K 96.2; RULER 128K 95.8 | vs BF16 | L82 @ main |
| r0b0tlab/EXL3-4.00bpw weights on disk 15.4 GiB [L31]4.00 bpw decoder, 6 bpw lm_head [L23] | EXL3 4.00 bpw | EXL3 | 16.54 | download | — | — | — | — | bpw 4.00 bpw | reference not stated | L31 @ main |
| Rabbit-Hole-Ai/5090-goldilocks-GGUF size 23.4 GiB from prose [L283]; bpw label 7.36 [L185]median KLD c42 also reported [L222] | imatrix build (thireus-7.4bpw-imatrix) | GGUF | 25.13 | download | 0.001678 | 98.278% | — | — | kld_99 0.014266 | vs BF16 | L185 @ main |
| Rabbit-Hole-Ai/5090-goldilocks-GGUF median KLD c38 [L221] | RTN build (thireus-7.4bpw, original) | GGUF | — | 0.001800 | 98.050% | — | — | kld_99 0.015519 | vs BF16 | L96 @ main | |
| Rabbit-Hole-Ai/5090-goldilocks-GGUF table of builds vs this cook [L185] | this cook (7.36 bpw class), corpus A | GGUF | — | — | — | — | — | kld:corpus A 0.001606; top1:corpus A 98.250% | vs BF16 | L185 @ main | |
| RadixArk/NVFP4-BF16-LMHead also evaluated on Terminal-Bench 2.1 per L89; only GSM8K in candidatescalibration: 1,024 cnn_dailymail samples, seq len 512 [L81] | Qwen3.8-27B-NVFP4-BF16-LMHead | NVFP4 | — | — | — | — | — | GSM8K 96.13% (1,268/1,319) | reference not stated | L123 @ main | |
| rapid-mlx/mixed-3.5bpw-MLX identical-subset comparison [L123-127] | mixed 3.5bpw (subset) | MLX | 13.00 | download | — | — | — | — | MMLU-Pro (n=200) 60.5; GSM8K (n=60) 93.3; MMBench-EN (n=150) 86.0 | other or no reference | L125 @ main |
| rdtand/PrismaAQUA-5.5bit-vllm max KL 1.3292 (c4.1) not mappedKL is truncated to teacher top-1024 plus lumped bucket, coverage mean 0.988 min 0.573 [L56-57]5.5002 bits per parameter excludes BF16 lm_head and embed_tokens [L32,L198]no downstream task suite run by the pipeline [L209] | PrismaQuant AQUA-AURA 5.5 bpp | other | 23.60 | download | — | — | — | — | kld:all_positions 0.0338; kld:confident_positions 0.0187; kld_99 0.2917 | vs BF16 | L32 @ main |
| rdtand/PrismaAQUA-gridbook-13GB-5080-vllm KL is a top-K logprob approximation, not full-vocab [L338]; mean_kld mapped to the all-positions value c35 [L323]size is weights+codebooks 12.959 GB (12.069 GiB) [L39]; bpw 3.855 over all 498 units, body-only 3.604 [L54-57]internal embed/head per-unit KL options [L243-248] are allocator choices, not mapped | PrismaAQUA gridbook 13GB | other | 12.96 | download | 0.09167 | 86.47 % | — | +4.56 | kld:confident_positions 0.05617; top1:confident_positions 96.87 %; ppl:wikitext2 9.792 | vs BF16 | L39 @ main |
| rdtand/PrismaScout-AQUA-Vision-20GB size 19978899058 bytes (18.6068 GiB) and 5.0737 bits per quantizable parameter are prose only [L50-53]KL confident-positions 0.0254206691 [c2]; KL max 2.379668877 [c4.1] | PrismaScout-AQUA-Vision-20GB | other | — | 0.0439023734 | — | 9.648316858 | — | kld_99 0.409348640 | vs BF16 | L115 @ main | |
| RedHatAI/MXFP4 no candidate table cells; scores given in prose list below [L143,L149,L155,L161]v2 rows (AIME 2026 97.8%, Terminal-Bench 2.1 55.4%) compare two checkpoints; attribution to this model not fully visible [L249,L260,L277-278]v1 scores are 3-seed (AIME25 8-seed) averages [L170] | Qwen3.8-27B-MXFP4 | MXFP4 | — | — | — | — | — | GSM8K Platinum 96.25%; MATH-500 83.67%; AIME 2025 96.67%; GPQA Diamond 89.56%; AIME 2026 97.8%; Terminal-Bench 2.1 55.4% | reference not stated | L143 @ main | |
| Reza2kn/Bev weights not trained or quantized by Bev; redistribution of Prism ML ternary model [L27]7.21 GB model file per headline [L21]benchmark is synthetic and correlated, not a quant fidelity measure [L120] | Bev (byte-identical Prism Ternary-Bonsai-2-27B GGUF) | other | 7.2 | download | — | — | — | — | Jev Noul % 95.00 | other or no reference | L21 @ main |
| RonnieOps/uncensored-IQ4_XS KL vs BF16 undefined for pruned vocab, not measured [L75,L81]standalone PPL 6.4248 +/- 0.04 in prose [L80]custom domain suite, not public benchmark [L167] | E4-ASCII pruned IQ4_XS (this quant) | other | 14.58 | download | — | — | 6.4248 | — | — | vs BF16 | L57 @ main |
| RukaRat/INT8-W8A8-imatrix-heretic-MTP noise floor about 2 items [L93-94]NER 28/40 vs 33-35 on other builds [L96-97]refusal rates: harmful 0% off c19, 3.3% on c20; benign-but-refused 0% on c21 [L67]trial-time Heretic KL 0.0621 is not deployed behaviour [L60-62] | INT8 W8A8 imatrix heretic MTP (this one) | INT8 | 31.25 | download | — | — | — | — | quality/160 130 | other or no reference | L48 @ main |
| RukaRat/W4A16-imatrix-heretic-int8head-MTP Size 16.6 GiB c18 in context table [L178]Card says 130-135/160 spread is a tie, rerun of same build moved 2 items [L226-228]Refusal evals: harmful 0/60 thinking off c6.0/c6.1, 1/30 thinking on c7, benign-but-refused 0/60 and 0/30 [L66-69]; identical output vs plain build 183/200 c2.0/c2.1 [L61] | W4A16 + int8 head, abliterated (this one) | INT8 | — | — | — | — | — | five_task_battery_correct_of_160 134 | other or no reference | L218 @ main | |
| sakamakismile/EXL3-6.5bpw same recipe baked three times with different md5 [L54]; 20-question probe, cannot rank [L34]; Huihui and Qwopus rows are other models, omitted [L29-31]published casting is the best seconds-per-correct of the three (value 27.0) [L56] | Qwen3.8-27B EXL3 6.5bpw casting a | EXL3 | — | — | — | — | — | MMLU-Pro_first20 15 | other or no reference | L28 @ main | |
| sakamakismile/EXL3-6.5bpw same recipe baked three times with different md5 [L54]; 20-question probe, cannot rank [L34]; Huihui and Qwopus rows are other models, omitted [L29-31] | Qwen3.8-27B EXL3 6.5bpw casting b | EXL3 | — | — | — | — | — | MMLU-Pro_first20 15 | other or no reference | L28 @ main | |
| sakamakismile/EXL3-6.5bpw same recipe baked three times with different md5 [L54]; 20-question probe, cannot rank [L34]; Huihui and Qwopus rows are other models, omitted [L29-31] | Qwen3.8-27B EXL3 6.5bpw casting c | EXL3 | — | — | — | — | — | MMLU-Pro_first20 16 | other or no reference | L28 @ main | |
| sammyblues/IQ2_XS-GGUF HellaSwag-100 intervals wide, directional only [L102-103] | 64L IQ2_XS (published) | GGUF | 9.1 | download | — | — | — | — | ppl:wikitext2 8.0306; HellaSwag-100 75% | vs BF16 | L108 @ main |
| sdkyuan/qat-q2_0-gguf table cells are HTML without candidate ids; values given via prose [L49-52] | Qwen3.8-27B-QAT-Q2_0 | GGUF | 8.8 | download | — | — | — | — | top1:reasoning 94.7%; kld:reasoning 0.029; top1:code 93.3%; kld:code 0.064; top1:tool_calling 78.7%; kld:tool_calling 0.638; top1:wikitext 78.3%; kld:wikitext 0.302 | other or no reference | L48 @ main |
| shisa-ai/FP8-BLOCK Median KLD 0.001737 c6 [L55]; harness resolves differences to ~1e-3, bootstrap CI +/-2e-3 [L50, L63]GDN in_proj_qkv/z/out_proj kept BF16 vs official FP8 [L39-40] | Qwen3.8-27B-FP8-BLOCK (this model) | FP8 | 36.40 | download | 0.004110 | 97.18% | — | — | kld_99 0.041333 | vs BF16 | L55 @ main |
| sigmanih/Qwen-GGUF-IQ1_S measured on a slice of each dataset, not full suite (7-14 questions each) [L100]tool-protocol adherence 62.0% c38, autonomous completion 0.0% c40 [L107-108] | Qwen3.8-27B GGUF IQ1_S | GGUF | 8.8 | download | — | — | — | — | ARC-Challenge 67%; GPQA 11%; GSM8K 33%; HellaSwag 44%; HumanEval 43%; MATH 0%; MBPP 33%; MMLU 36%; MMLU-Pro 11%; TruthfulQA 67% | reference not stated | L46 @ main |
| sigmanih/Qwen-GGUF-IQ2_S measured on a slice of each dataset, not the full suite; per-benchmark n is 7-14 questions [L100,L85-96]Autonomous task completion 0.0% (0/1) not mapped [L108]Thinking mode 28/43 and no-thinking 43/57 [L78-79] | Qwen3.8-27B-GGUF-IQ2_S | GGUF | 10.71 | download | — | — | — | — | Overall 71%; ARC-Challenge 100%; BIG-Bench Hard 86%; GPQA 22%; GSM8K 89%; HellaSwag 78%; HumanEval 100%; MATH 67%; MBPP 56%; MMLU 57%; MMLU-Pro 67%; TruthfulQA 78% | other or no reference | L46 @ main |
| sigmanih/Qwen-GGUF-IQ2_XXS measured on a slice of each dataset, not full suite; not comparable with full-suite runs [L100]per-dataset n is 7-14 questions [L85-96]no-thinking 80.7% (c8) vs thinking 67.4% (c10) [L78-79]tool protocol 75.0% adherence, autonomous completion 0.0% [L107-108] | IQ2_XXS | GGUF | 9.9 | download | — | — | — | — | ARC-Challenge 100%; BIG-Bench Hard 100%; GPQA 22%; GSM8K 89%; HellaSwag 67%; HumanEval 100%; MATH 78%; MBPP 89%; MMLU 64%; MMLU-Pro 56%; TruthfulQA 78% | other or no reference | L46 @ main |
| sigmanih/Qwen-GGUF-Q4_K_M Measured on a slice of the dataset, not full suite; not comparable with full-suite runs [L100]Tiny samples (7-14 questions per dataset) [L85-96]Overall 79/100; no-thinking 77.2% (44/57), thinking 81.4% (35/43) [L78-79]Tool protocol adherence 94.6% but listed as 0/92 criteria verified and autonomous task completion 0/1 [L107-108] | Q4_K_M GGUF | GGUF | 15.66 | download | — | — | — | — | Overall 79%; ARC-Challenge 100%; BIG-Bench Hard 86%; GPQA 56%; GSM8K 100%; HellaSwag 67%; HumanEval 100%; MATH 89%; MBPP 67%; MMLU 57%; MMLU-Pro 78%; TruthfulQA 89% | reference not stated | L46 @ main |
| sigmanih/Qwen-GGUF-Q5_K_M measured on a slice of datasets (75/100 questions overall, 7-14 per dataset), not comparable to full-suite runs [L100]tool-protocol adherence 93.5% but 0/92 criteria and 0/1 scenarios completed [L107-108] | Qwen3.8-27B-GGUF-Q5_K_M | GGUF | 18.19 | download | — | — | — | — | ARC-Challenge 100%; BIG-Bench Hard 71%; GPQA 22%; GSM8K 89%; HellaSwag 67%; HumanEval 100%; MATH 89%; MBPP 78%; MMLU 64%; MMLU-Pro 67%; TruthfulQA 89%; NoThinking 80.7%; Thinking 67.4% | other or no reference | L46 @ main |
| signalnine/MTP-q27 HumanEval+ is 29/30; id gives solved count out of 30 [L34] | q6k (v2) | other | 22.52 | download | — | — | 7.1718 ppl:wikitext | — | HumanEval+ 29 | reference not stated | L34 @ main |
| signalnine/MTP-q27 HumanEval+ is 28/30; id gives solved count out of 30 [L33] | q6 (v2) | other | 19.76 | download | — | — | 7.2233 ppl:wikitext | — | HumanEval+ 28 | reference not stated | L33 @ main |
| signalnine/MTP-q27 HumanEval+ is 30/30; id gives solved count out of 30 [L32] | default (v2) | other | 17.00 | download | — | — | 7.3121 ppl:wikitext | — | HumanEval+ 30 | reference not stated | L32 @ main |
| signalnine/MTP-q27 HumanEval+ is 30/30; id gives solved count out of 30 [L31] | q4s (v2) | other | 15.70 | download | — | — | 7.3765 ppl:wikitext | — | HumanEval+ 30 | reference not stated | L31 @ main |
| singulared/ROCmFP4-MTP-GGUF PPL +/- 0.0417; +0.129 vs STRIX (c54) [L147] | ROCMFP4-COHERENT | GGUF | 15.47 | download | — | — | 6.5002 ± 0.0417 | — | — | other or no reference | L147 @ main |
| singulared/ROCmFP4-MTP-GGUF PPL +/- 0.0402 [L145]vs-STRIX delta column not mapped (not a listed key) [L142-147] | ROCMFP4-STRIX | GGUF | 14.76 | download | — | — | 6.3715 ± 0.0402 | — | — | other or no reference | L145 @ main |
| singulared/ROCmFP4-MTP-GGUF PPL +/- 0.0404; +0.025 vs STRIX (c51) [L146]card says FAST is 4.25 bpw [L218] | ROCMFP4-FAST | GGUF | 14.31 | download | — | — | 6.3968 ± 0.0404 | — | — | other or no reference | L146 @ main |
| slevinw/Heretic-Abliterated-Uncensored-GGUF delta vs F16 [L345] | RVN-Q5_K_M.gguf (standard) | GGUF | — | — | — | — | +2.23% | ppl:tiny_shakespeare 4.6493 | other or no reference | L345 @ main | |
| slevinw/Heretic-Abliterated-Uncensored-GGUF text perplexity only; vision not measured [L346,L350] | RVN-Q5_K_M-vision.gguf | GGUF | — | — | — | — | +2.24% | ppl:tiny_shakespeare 4.6497 | other or no reference | L346 @ main | |
| slevinw/Heretic-Abliterated-Uncensored-GGUF [L347] | RVN-Q4_K_M-vision.gguf | GGUF | — | — | — | — | +7.20% | ppl:tiny_shakespeare 4.8751 | other or no reference | L347 @ main | |
| slevinw/Heretic-Abliterated-Uncensored-GGUF [L348] | RVN-Q3_K_M-vision.gguf | GGUF | — | — | — | — | +24.2% | ppl:tiny_shakespeare 5.6490 | other or no reference | L348 @ main | |
| slevinw/Heretic-Abliterated-Uncensored-GGUF KL is abliteration damage vs base model, not quantization KL; refusals 0-1/100 range on 100 harmful prompts [L128-132] | RVN (this repo, abliteration vs base) | GGUF | — | 0.0085 | — | — | — | — | other or no reference | L132 @ main | |
| slevinw/Heretic-Abliterated-Uncensored-GGUF absolute PPL deltas, eval corpus not stated [L239-243] | Q4_K_M multilingual (pilot) | GGUF | — | — | — | — | — | ppl_delta:wiki_imatrix_vs_F16 -0.0002; ppl_delta:wiki_imatrix_vs_legacy +0.0156; ppl_delta:multilingual_vs_F16 +0.0038; ppl_delta:multilingual_vs_legacy -0.0074 | other or no reference | L243 @ main | |
| slevinw/Heretic-Abliterated-Uncensored-GGUF absolute PPL deltas, eval corpus not stated [L239-244] | IQ4_XS multilingual (pilot) | GGUF | — | — | — | — | — | ppl_delta:wiki_imatrix_vs_F16 +0.0290; ppl_delta:wiki_imatrix_vs_legacy +0.0060; ppl_delta:multilingual_vs_F16 +0.0091; ppl_delta:multilingual_vs_legacy -0.0083 | other or no reference | L244 @ main | |
| SmokeyPete/Heretic-NVFP4-DGX-Spark calibration abisee/cnn_dailymail 1,024 x 512 tokens [L205]HumanEval base BF16 97.56% given but this-model value not in candidates [L114] | Heretic NVFP4 (this model) | NVFP4 | 21.95 | download | — | — | — | — | GSM8K 96.82%; HumanEval+ 92.68% | vs BF16 | L41 @ main |
| soyrsoyr/W4A16-AWQ-GPTQ recovery % vs bf16: c3,c6,c9,c12,c15,c18,c21,c24,c27 [L81-100]MMLU and GSM8K omitted as harness artifacts [L87-90] | W4A16-AWQ-GPTQ | AWQ | — | — | — | — | — | OpenLLM v1 average 64.75; Reasoning average 85.54 | vs BF16 | L81 @ main | |
| staccs/lecore-mlx Card admits the export is uniform 4-bit group 64 with 4-bit embeddings and 6-bit lm_head, not a mix; fair baseline is stock MLX 4-bit [L20]Benchmarks are subsets, about +-2-4 points noise [L34] | leCore Unicron mix (this repo) | MLX | 15.45 | text weights (excl. vision, MTP) | — | — | — | — | ppl:wikitext2 8.1249; ppl:python-code 2.7038; MMLU 5-shot 84.9%; GSM8K strict 91.0%; GSM8K flexible 89.0% | vs BF16 | L30 @ main |
| Steven10429/UD-IQ4_XS-MLX size on disk 16.45 GB is in the row label of [L32], no candidate idre-quantization of a quantized model [L113] | this repo (UD-IQ4_XS to MLX) | MLX | — | — | — | — | — | ppl:wikitext2 6.4428 | vs BF16 | L33 @ main | |
| stosc/qwen3_8_27b_nvfp4_v100.ninfer IFBench is prompt-level strict [L264]; instruction-level strict 80.50%, prompt-level loose 80.33%, instruction-level loose 83.50% [L271]File size 22.09 GiB (23,719,715,844 bytes) is in a separate artifact table [L119] | NInfer NVFP4 | NVFP4 | — | — | — | — | — | IFBench 77.00%; AIME 2025 96.67%; AIME 2026 96.67%; GPQA-Diamond 90.40%; ERQA 66.25%; RealWorldQA 83.53% | vs BF16 | L264 @ main | |
| sunnyyy/NInfer IFBench is prompt-level strict [L187]deltas within +/-3.1 points on four overlapping benchmarks [L201] | NInfer groupwise-int | other | — | — | — | — | — | AIME2025 96.67%; AIME2026 96.67%; GPQA-Diamond 87.37%; ERQA 66.25%; RealWorldQA 82.22% | vs BF16 | L187 @ main | |
| sunnyyy/nvfp4-NInfer IFBench is prompt-level strict [L273]accuracy, rule-based scoring [L265-267] | Qwen3.8-27B NVFP4 (NInfer) | NVFP4 | — | — | — | — | — | AIME 2025 96.67%; AIME 2026 96.67%; GPQA-Diamond 90.40%; ERQA 66.25%; RealWorldQA 83.53% | other or no reference | L273 @ main | |
| syvai/3090-fast-variant Extras repo: only lm_head and MTP draft requantized to int4, perplexity +0.6% vs bf16, GSM8K unchanged [L13] | model-00007-of-00007.safetensors (lm_head requantized to int4) | other | — | — | — | — | +0.6% perplexity vs bf16 | — | vs BF16 | L13 @ main | |
| talxcc/Tals-coder-flash-01 retention % vs Q4_K_M with no per-benchmark scores or method [L47-48,65] | Tals-coder-flash-01 | other | 11.86 | download | — | — | — | — | Coding_Retention 99.9% | other or no reference | L65 @ main |
| tcclaviger/RFI8-AA as shipped [L41]Qwen3.8-27B upstream benchmark tables (L135-285) are the base model's, omitted | W8A8, fp8 KV cache (as shipped) | other | — | — | — | 6.9392 | — | — | vs BF16 | L41 @ main | |
| TechnoBaptist/Ternary-Bonsai-2-27B-gguf Card text appears to mirror the PrismML Bonsai 2 27B card; benchmark figures are as printed [L31-35,L206-256]GSM8K cell for Bonsai missing from per-benchmark table [L244]MBPP+ missing for UD-Q4_K_XL and MMMU-Pro missing for UD-Q4_K_XL in table [L249,L254]Sizes describe language model alone, vision tower separate [L85] | Bonsai 2 27B | GGUF | 5.9 | text weights (excl. vision, MTP) | — | — | — | — | MMLU-Redux 89.09; MuSR 70.63; MATH-500 98.80; AIME25 95.00; AIME26 95.83; HumanEval+ 95.12; MBPP+ 83.07; LiveCodeBench 90.07; IFEval 91.31; IFBench 74.00; BFCL v3 74.92; MMMU-Pro 75.49; OCR Bench v2 56.88 | other or no reference | L215 @ main |
| TelperionAI/EXL3-4.0bpw [L34]near-tie/moderate/certain/divmed columns and KV-cache table (L132-134) not mappedconfident is the pre-registered decider, lower is better [L126] | EXL3 4.0 bpw (this model, AWQ-smoothed) | EXL3 | 16.00 | download | — | 93.82% | — | — | — | vs BF16 | L34 @ main |
| TelperionAI/EXL3-4.0bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38] | TelperionAI EXL3 4.0bpw-v2 | EXL3 | 16.00 | download | — | 95.16% | — | — | — | vs BF16 | L43 @ main |
| TelperionAI/EXL3-5.5bpw [L31]near-tie/moderate/certain/divmed columns and KV-cache table (L132-134) not mappedconfident is the pre-registered decider, lower is better [L126] | EXL3 5.5 bpw (this model, AWQ-smoothed) | EXL3 | 20.00 | download | — | 96.12% | — | — | — | vs BF16 | L31 @ main |
| TelperionAI/EXL3-5.5bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38] | TelperionAI EXL3 5.5bpw-v2 | EXL3 | 20.00 | download | — | 96.26% | — | — | — | vs BF16 | L43 @ main |
| TelperionAI/EXL3-6.5bpw [L30]near-tie/moderate/certain/divmed columns and KV-cache table (L132-134) not mappedconfident is the pre-registered decider, lower is better [L126] | EXL3 6.5 bpw (this model, AWQ-smoothed) | EXL3 | 23.00 | download | — | 97.10% | — | — | — | vs BF16 | L30 @ main |
| TelperionAI/FP8-block-AWQ Top-1 column read as agreement with BF16 base [L28-30]Near-tie/moderate/confident/certain disagreement buckets, divmed and tok/s not mapped (no key) [L30] | FP8 block + AWQ (this model) | AWQ | 30.00 | unknown | — | 96.22% | — | — | — | vs BF16 | L30 @ main |
| TelperionAI/INT4-AWQ-GPTQ size includes ~0.85 GB BF16 MTP head [L61]perplexity deliberately excluded [L71]bucket columns are disagreement rates; divmed higher is better | this model (INT4 AWQ+GPTQ) | AWQ | 25.10 | download | — | 96.30% | — | — | — | vs BF16 | L55 @ main |
| TelperionAI/INT4-AWQ-GPTQ-gdn4 top-1 is argmax agreement with BF16 [L66] | this model (GDN 4-bit) | AWQ | 22.60 | download | — | 95.88% | — | — | near-tie disagree 25.26%; moderate disagree 3.60%; confident disagree 1.15%; certain disagree 0.13%; divmed 51 | vs BF16 | L59 @ main |
| TelperionAI/NVFP4-AWQ-AutoRound size includes ~0.85 GB BF16 MTP head [L63]no KLD or perplexity reported [L73][L57] | this model | NVFP4 | 24.70 | download | — | 93.38% | — | — | divmed 28 | vs BF16 | L57 @ main |
| TelperionAI/NVFP4-AWQ-AutoRound-allfp4 size includes ~0.85 GB BF16 MTP head [L64]no KLD or perplexity reported; perplexity excluded on purpose [L74][L57] | this model (all-FP4) | NVFP4 | 23.80 | download | — | 93.05% | — | — | — | vs BF16 | L57 @ main |
| TelperionAI/NVFP4-AWQ-GPTQ size includes ~0.85 GB BF16 MTP head [L63]perplexity deliberately excluded [L74] | this model (NVFP4+AWQ) | NVFP4 | 24.70 | download | — | 93.44% | — | — | — | vs BF16 | L57 @ main |
| ThakiCloud/NVFP4-FP8ATTN GSM8K strict delta vs bf16 -8.33pp, p 0.004 [L92]Rebuild noise floor for GSM8K strict is 3.56pp [L117] | this build (NVFP4 MLP + FP8 attention) | NVFP4 | 22.90 | download | — | — | — | — | MMMU-val MC 0.8707; HAERAE 0.7654; KoBEST 0.6670; belebele-ko 0.8983; ARC-Challenge 0.5733; IFEval prompt-strict 0.3179 | vs BF16 | L45 @ main |
| ThakiCloud/NVFP4-GPTQ-mm +0.8pp vs bf16 c3, McNemar p = 0.754 c4 [L107]; no detectable quality differenceGSM8K not run on this build [L134]Earlier bf16 MMMU 0.4511 withdrawn as harness error [L115] | this build (Qwen3.8-27B-NVFP4-GPTQ-mm) | NVFP4 | 30.14 | download | — | — | — | — | MMMU-val MC 0.8537 | vs BF16 | L36 @ main |
| ThakiCloud/NVFP4-GPTQ-txt GSM8K gap vs bf16 is within the author's rebuild-to-rebuild noise floor, so not attributable [L231-240]MMMU McNemar p = 0.453, no detectable difference [L107,L109]Only MLP stack is 4-bit; attention and others kept bf16 [L34-53] | this build (NVFP4-GPTQ-txt) | NVFP4 | 30.14 | download | — | — | — | — | MMMU-val MC 0.8577; HAERAE 0.7608; KoBEST 0.6745; belebele-ko 0.9083; ARC-Challenge 0.5750; IFEval prompt-strict 0.3142 | vs BF16 | L36 @ main |
| ThakiCloud/W4A16-AWQ card says score with flexible-extract, not strict-match [L18]size from prose [L13]one run, GSM8K only [L86] | this build (AWQ) | AWQ | 18.59 | download | — | — | — | — | GSM8K_en_strict 0.4147; GSM8K_en_flexible 0.5299; GSM8K_ko_strict 0.5921; GSM8K_ko_flexible 0.6156 | vs BF16 | L13 @ main |
| ThakiCloud/W4A16-GPTQ Scores printed as fractions [L40]Rebuild noise floor: two identical-recipe builds differ by 3.56pp on GSM8K en strict, 27.08pp on RAG fabrication rate [L51-65]RAG fabrication 9.17% (n=240) for this build not mapped [L64] | W4A16 GPTQ (this build) | GPTQ | 18.59 | download | — | — | — | — | GSM8K-en-strict 0.5155; GSM8K-en-flexible 0.5603; GSM8K-ko-strict 0.5785; GSM8K-ko-flexible 0.6505 | vs BF16 | L40 @ main |
| thederby/oQ4e-fp16-mtp this repo (bold row) [L27] | oQ4e fp16 | other | 17.90 | download | 0.0351 | 90.5% | 10.47 | — | — | vs BF16 | L27 @ main |
| thederby/oQ4e-mtp this repo, bold row [L26] | oQ4e bf16 | other | 17.00 | download | 0.0358 | 90.6% | 10.47 | — | — | vs BF16 | L26 @ main |
| thederby/oQ5e-fp16-mtp this repo in bold [L22]fp16 builds ~0.9 GB larger because vision tower kept in float32 [L38-39] | oQ5e fp16 | other | 21.20 | download | 0.0103 | 94.9% | 10.28 | — | — | vs BF16 | L29 @ main |
| thederby/oQ5e-mtp bold row = this repo [L28]KL slightly underestimates full KL [L44] | oQ5e bf16 | other | 20.30 | download | 0.0106 | 94.9% | 10.28 | — | — | vs BF16 | L28 @ main |
| thederby/oQ6e-fp16-mtp this repo in bold [L31] | oQ6e fp16 | other | 24.60 | download | 0.0028 | 97.1% | 10.22 | — | — | vs BF16 | L31 @ main |
| thederby/oQ6e-mtp this repo in bold [L22]top-256 truncated KL [L44] | oQ6e bf16 | other | 23.70 | download | 0.0033 | 97.2% | 10.23 | — | — | vs BF16 | L30 @ main |
| thederby/oQ8e-fp16-mtp this repo, bold row [L33]KL computed on top-256 tokens, compare only with same method [L43-45] | oQ8e fp16 | other | 30.90 | download | 0.0007 | 98.4% | 10.20 | — | — | vs BF16 | L33 @ main |
| thederby/oQ8e-mtp this repo in bold [L32] | oQ8e bf16 | other | 30.00 | download | 0.0010 | 98.6% | 10.21 | — | — | vs BF16 | L32 @ main |
| TheDrainFlorist/VQ-3.9bpw superseded table [L121] | this model [old prose-only 2048 table] | other | 13.39 | download | — | — | 5.229 | — | kl_mnats:prose 85.8; top1:prose 86.1% | vs BF16 | L121 @ main |
| TheDrainFlorist/VQ-4.8bpw superseded, not comparable with three-corpus table [L111,L122] | VQ-4.8bpw (superseded prose-only 2048 table) | other | 16.59 | download | 32.8 mnats | 90.8% | 5.162 | — | — | vs BF16 | L54 @ main |
| thousandml/CQ-4bit [L36,L37]calibration-free 4-bit method [L26] | CQ 4-bit | other | — | 0.0282 | 92.4% | — | — | — | other or no reference | L36 @ main | |
| timteh673/Opus-Abliterix-Reasoning-MLX-8bit table compares BF16 models, not the MLX 8-bit release itself [L39]Benign KL is divergence from control, an abliteration effect, not quantization KL [L52]Full code count is passes out of 421 [L47]held-out loss ratio omitted as ambiguous [L51] | abliterix-pass1-bf16 (winner) | MLX | — | 0.093614 | — | — | — | Capability macro 21.0086%; Full code 10; HumanEval 4.2683%; Long-form pass 62.5000%; MMMU30 11; Harmful hard refusal 0.0% | vs BF16 | L43 @ main | |
| ToBeStyled/ColdFusion-GAIN-Blackwell-DFlash2-Ultra-V1.0 perplexity 2.8059 +/- 0.034 [L112]+15.5% decode for +2.7% perplexity vs Q4_K_M [L115] | this repo NVFP4 + Q8_0 head/embed | other | 16.90 | download | — | — | 2.8059 ± 0.034 | — | — | other or no reference | L112 @ main |
| ToBeStyled/ColdFusion-GAIN-Blackwell-DFlash2-Ultra-V1.0 build used for accuracy table not specified [L65-68] | This model (accuracy table) | other | — | — | — | — | — | HumanEval-50 pass@1 94.0%; GSM8K-50 96.0% | other or no reference | L67 @ main | |
| tooltd/GSQ-RCO-BuffedMod-GGUF identical core metrics to IQ4XS MTP variant [L82] | BuffedMod (Q6_K MTP, Q6_K output.weight) | GGUF | 11.59 | download | 0.0517 | 90.54% | 7.0321 | — | kld_99 KLD: 0.586696 | vs BF16 | L78 @ main |
| tooltd/GSQ-RCO-BuffedMod-GGUF detailed stats L115-135 shared by both BuffedMod rows; table values identical for both [L77-78] | BuffedMod (IQ4XS MTP, Q6_K output.weight) | GGUF | 11.49 | download | 0.0517 | 90.54% | 7.0321 | — | kld_99 KLD: 0.586696 | vs BF16 | L77 @ main |
| tooltd/IQ4-XS-16GB-VRAM-GGUF Same Top-p (%) column mapped to top1_pct; label ambiguous [L58]EvalPlus pass@1 with q4_0 KV cache [L32] | ZB4.00-MIN-v5.1-IQ4_XS | GGUF | 12.74 | download | 0.033810 | 92.309% | 7.106519 | — | HumanEval 0.945; HumanEval+ 0.921; MBPP 0.897; MBPP+ 0.780 | vs BF16 | L37 @ main |
| tooltd/IQ4-XS-16GB-VRAM-GGUF | ZB3.88-MIN-v5-IQ3_M_XL | GGUF | 12.34 | download | 0.042164 | 91.452% | 7.132594 | — | HumanEval 0.945; HumanEval+ 0.915; MBPP 0.91; MBPP+ 0.775 | vs BF16 | L39 @ main |
| tooltd/IQ4-XS-16GB-VRAM-GGUF | ZB3.73-MIN-v5.1-IQ3_M_L | GGUF | 11.88 | download | 0.048117 | 90.759% | 7.199937 | — | HumanEval 0.927; HumanEval+ 0.896; MBPP 0.881; MBPP+ 0.757 | vs BF16 | L43 @ main |
| tooltd/IQ4-XS-16GB-VRAM-GGUF | GSQ-RCO-BuffedMod | GGUF | 11.49 | download | 0.051738 | 90.535% | 7.032142 | — | HumanEval 0.921; HumanEval+ 0.890; MBPP 0.897; MBPP+ 0.772 | vs BF16 | L41 @ main |
| tooltd/IQ4-XS-16GB-VRAM-GGUF EvalPlus row labelled ZB3.00bpw-IQ3_XXS [L45] | ZB3.0BPW-IQ3_XXS | GGUF | 9.6 | download | 0.120503 | 85.320% | 7.474669 | — | HumanEval 0.927; HumanEval+ 0.884; MBPP 0.865; MBPP+ 0.743 | vs BF16 | L45 @ main |
| trevest/Tetris-GGUF ppl_delta_pct is vs Q8_0 not BF16 [L24-28] | Tetris 4.4bpw | GGUF | 15.06 | download | — | — | 7.1453 | +2.24% | — | vs a Q8_0 | L28 @ main |
| Triody/NVFP4-AllLinear two shards 9,959,955,840 + 7,034,658,960 bytes, no single size id [L126-127]speed table (tokens/s) not mapped [L95-99] | Qwen3.8-27B-NVFP4-AllLinear (int8 draft) | NVFP4 | — | — | — | — | — | GSM8K_lenient 97.25%; GSM8K_strict 96.00%; HumanEval_lenient 91.46%; HumanEval_strict 86.59% | vs BF16 | L110 @ main | |
| UgurI/Turkish-CPT-GGUF Continued-pretraining fine-tune, so deltas vs base reflect Turkish CPT, not quantization loss [L19, L36-42]Q4_K_M 15.41 GB, 4.92 BPW stated [L24] | Qwen3.8-27B-Turkish-CPT-Q4_K_M.gguf | GGUF | 15.41 | download | — | — | — | — | ppl:turkish_validation 6.9360; ppl:turkish_test 7.3303; TurkishMMLU 75.56%; Belebele TR 90.00%; Belebele EN 96.67% | other or no reference | L24 @ main |
| varshavvvvv/Ternary-CyberX-2-27B-gguf ternary; same quality numbers apply to both PTQ1_0 and PQ2_0 packings, which are not benchmarked separately [L44,L85]GSM8K has no Bonsai cell [L244] | Bonsai 2 27B | GGUF | 5.9 | text weights (excl. vision, MTP) | — | — | — | — | MMLU-Redux 89.09; MuSR 70.63; MATH-500 98.80; AIME25 95.00; AIME26 95.83; HumanEval+ 95.12; MBPP+ 83.07; LiveCodeBench 90.07; IFEval 91.31; BFCL v3 74.92; MMMU-Pro 75.49; OCR Bench v2 56.88 | other or no reference | L215 @ main |
| vrfai/FP8 Drift table [L99-110] reports delta vs bf16 and noise floor per task (AI2D, ChartQA, ScienceQA-IMG, MMStar, OCRBench, CountBenchQA, RealWorldQA, OmniDocBench, MathVision, ERQA); deltas not mapped, no matching metric keySize 28.3 GB is the Quantized weights row [L124]; base model 52 GB [L23] | Qwen3.8-27B-FP8 (This model) | FP8 | 28.30 | download | — | — | — | — | MathVision 84.54; ERQA 63.75 | vs BF16 | L75 @ main |
| vrfai/FP8-dynamic absolute scores, MathVision n=304 ERQA n=400 [L75-76]size 28.3 GB vs base 52 GB [L23] | Qwen3.8-27B-FP8-dynamic (absolute scores) | FP8 | 28.30 | download | — | — | — | — | MathVision 82.89; ERQA 65.25 | vs BF16 | L75 @ main |
| vrfai/FP8-dynamic deltas in points vs bf16; noise floors listed per task (c50,c53,c56,c59,c62,c65,c68,c71,c74,c77) [L101-110]deltas smaller than floor are measurement noise per card [L112-113] | Qwen3.8-27B-FP8-dynamic (drift vs bf16) | FP8 | — | — | — | — | — | AI2D_delta_vs_bf16 -0.58; ChartQA_delta_vs_bf16 -0.08; ScienceQA-IMG_delta_vs_bf16 +0.00; MMStar_delta_vs_bf16 +2.33; OCRBench_delta_vs_bf16 +0.50; CountBenchQA_delta_vs_bf16 +0.00; RealWorldQA_delta_vs_bf16 +0.26; OmniDocBench_delta_vs_bf16 -2.26; MathVision_delta_vs_bf16 -3.29; ERQA_delta_vs_bf16 +0.25 | vs BF16 | L101 @ main | |
| vrfai/NVFP4 Absolute scores only for MathVision and ERQA [L79-83]; other tasks are drift vs bf16 only, with a bf16-vs-bf16 noise floor [L90-105]Measured on H100 where activations stay bf16, so effectively W4A16 not W4A4; card calls these an upper bound on Blackwell accuracy [L29]Drift table verdict column not captured; deltas smaller than floor are noise per card [L118-123] | Qwen3.8-27B-NVFP4 (this model) | NVFP4 | 18.40 | download | — | — | — | — | MathVision 81.91; ERQA 66.25 | vs BF16 | L81 @ main |
| WarlaxZ/NInfer-3090 KLD on held-out corpus vs Unsloth Q8_0 [L83-90]Perplexity change vs upstream -0.93% / -0.50% not mapped [L80-81] | NInfer-3090 artifact (4-bit embedding) | other | — | 0.0318 | — | — | — | ppl:perplexity-1m 4.3023; ppl:heldout-2026-09 4.3178 | other or no reference | L80 @ main | |
| WaveCut/MLX-4bit-DWQ KL is against the 8-bit teacher, not BF16 [L45-L46]acceptance threshold and 63.596% reduction vs baseline stated in prose [L58-L59] | DWQ final (1,024 examples) | MLX | — | 0.074299 | — | — | — | — | other or no reference | L56 @ main | |
| wiklif/KTopt-GGUF Benchmark table header gives 10.3 GB [L65]; size from KLD table, GGUF + 32 MB adapter [L115]Benchmarks at 200k output limit with q8_0 KV [L63,L101] | KTopt + EoRA r64 | GGUF | 10.31 | download | 0.0644 kld:prose | — | — | — | GPQA Diamond 89.4 ± 4.3; IFBench_loose 85.3; IFBench_strict 81.7; LiveCodeBench v6 85.5 ± 6.0; kld:code 0.0419 | vs a Q8_0 | L67 @ main |
| wiklif/KTopt-GGUF without EoRA adapter [L116] | KTopt | GGUF | 10.31 | download | 0.0662 kld:prose | — | — | — | kld:code 0.0429 | vs a Q8_0 | L116 @ main |
| XpressAI/RYS-GGUF RYS value missing for simple_python, live_parallel, live_simple (change 0.00) [L73,L76,L79]BFCL gain not statistically decisive; math probe diagnostic only [L32-39]not the full 3,641-case BFCL suite [L62-63] | RYS 16-19 UD-Q4_K_XL | GGUF | 18.90 | download | — | — | — | — | Reasoning probe 70.59%; EQ-140 81.66; BFCL micro 83.20%; BFCL macro 81.81%; Math probe 0.750 | other or no reference | L43 @ main |
| yeasah/exl3 mean_kld is excess KLD [L40] | 6.00 bpw | EXL3 | 22.95 | download | — | — | — | +0.03% | — | vs BF16 | L33 @ main |
| yeasah/exl3 mean_kld is excess KLD [L40] | 5.00 bpw | EXL3 | 19.91 | download | — | — | — | +0.14% | — | vs BF16 | L32 @ main |
| yeasah/exl3 mean_kld is excess KLD [L40] | 4.00 bpw | EXL3 | 16.87 | download | — | — | — | +0.37% | — | vs BF16 | L31 @ main |
| yeasah/exl3 mean_kld is excess KLD [L40] | 3.00 bpw | EXL3 | 13.67 | download | — | — | — | +1.79% | — | vs BF16 | L30 @ main |
| yeasah/exl3 mean_kld is excess KLD, not raw KLD [L40]VRAM columns exllamav3 7.32 GiB, vLLM 7.13 GiB not used [L29] | 2.00 bpw | EXL3 | 10.63 | download | — | — | — | +9.62% | — | vs BF16 | L29 @ main |
| YourHighnessLA/L0xRE-27b-Low-MTP internal suite, directional only [L73]; three passes [L71] | L0xRE-27b-Low-MTP | other | — | — | — | — | — | internal_suite_pass@1 124–126 / 150 | other or no reference | L71 @ main | |
| Youssofal/MTPLX-Bare-Speed download size 16.0 GB [L96]; peak unified memory 17.0 GB [L97]KL 0.0376 on coding battery; different corpus than the 2,389-position figures [L90] | Bare Speed | other | 16.00 | download | — | — | — | — | kld:coding battery 0.0376 | vs BF16 | L96 @ main |
| Youssofal/MTPLX-Optimized-Quality Download size 29.4 GB; peak unified memory 32.7 GB c6 measured [L94-95]Card also says KL to bf16 on coding battery 0.00105, 21x closer than Optimized Speed and 36x closer than Bare Speed [L88-89] | Optimized Quality (8-bit dynamic) | other | 29.40 | download | — | — | — | — | top1:code_prose_json_multilingual 99.3 percent; kld:coding_battery 0.00105 | vs BF16 | L94 @ main |
| Youssofal/MTPLX-Optimized-Speed size is download [L98]peak unified memory 23.6 GB [L99]KL 0.0220 on coding battery [L93] is a different measurement | Optimized Speed (4-bit dynamic) | other | 20.40 | download | 0.012 | 96.0 percent top-1 agreement | — | — | kld:code 0.0220 | vs BF16 | L58 @ main |
| YTan2000/TQ | TQ_L | other | 14.72 | download | 0.022 | 93.7 % | — | — | GSM8K 94.9 %; GPQA-Diamond 87.88 %; BenchLoop_speed 91.9; BenchLoop_coding 87.5; BenchLoop_dataextract 88.4; BenchLoop_instructfollow 84.5 | vs BF16 | L85 @ main |
| YTan2000/TQ TQ_M has KLD only [L94] | TQ_M | other | 13.14 | download | 0.045 | 91.2 % | — | — | — | vs BF16 | L85 @ main |
| YTan2000/TQ GPQA on RTX 3090 [L92] | TQ_S | other | 12.55 | download | 0.055 | 90.1 % | — | — | GSM8K 94.6 %; GPQA-Diamond 82.83 %; BenchLoop_speed 93.2; BenchLoop_coding 85.4; BenchLoop_dataextract 89.0; BenchLoop_instructfollow 91.1 | vs BF16 | L85 @ main |
| YTan2000/TQ GSM8K and MMLU-Pro XS cells marked * were run on RTX 3090, others on GB10 [L88-L92]MMLU-Pro 291 questions has standard error about 2.4 points; GSM8K differences within noise [L93]BenchLoop on RTX 3090 [L96] | TQ_XS | other | 11.85 | download | 0.079 | 88.6 % | — | — | GSM8K 95.4 % \*; GPQA-Diamond 84.85 %; BenchLoop_speed 92.6; BenchLoop_toolcall 86.7; BenchLoop_coding 79.2; BenchLoop_dataextract 92.9; BenchLoop_instructfollow 76.7; BenchLoop_reasonmath 80.0 | vs BF16 | L85 @ main |
| YTan2000/TQ3_4S first-pass battery; reasoning-off [L167]tool-call 86.7 uses v2 template; v1 81.7 in prose [L52] | Qwen3.8-27B-TQ3_4S-v2.gguf (recommended) | other | 13.80 | download | — | — | — | — | Hard86 74/86 (86.0%); HumanEval 92.7; HumanEval+ 88.4; MBPP 90.5; MBPP+ 77.2; CodingSuite 93.8; ReasonMathSuite 86.7; ToolCallSuite 86.7; DataExtractSuite 80.2; InstructionFollowSuite 62.2 | other or no reference | L41 @ main |
| zarigata/FeveQwen-1.0 behavioral derivative (refusal-subspace projection), not a quant; KL/PPL reference and corpus not stated [L56-59]Ollama build is Q4_K_M at 4.92 bits/weight but metrics are not tied to it [L98] | FeveQwen 1.0 (tune-set candidate E02) | other | — | 0.106 | — | 3.42 | — | Coherence 100% (10/10); Capability_checks 83% (5/6) | reference not stated | L56 @ main | |
| zedongpeng/NInfer-4090 KL64 is top-64 truncated KL, not full-vocab [L49]top-1 printed as fraction 0.9250, not percent [L54]vision, MTP, long-context not evaluated [L57-58] | this model | other | — | 0.0619 | 0.9250 | 3.213 | — | kld:prose 0.0265; kld:code 0.1018 | vs BF16 | L54 @ main | |
| zerodigest/YMQ-MTP-GGUF size only in row label (~9.8 GB), no candidate id [L64] | XXS | GGUF | — | 0.193804 ± 0.0022 | — | — | — | ppl:wikitext2 7.7848 | reference not stated | L64 @ main | |
| zerodigest/YMQ-MTP-GGUF size only in row label (~10.2 GB) [L65] | XS-TI | GGUF | — | 0.168378 ± 0.001935 | — | — | — | ppl:wikitext2 7.6090 | reference not stated | L65 @ main | |
| zerodigest/YMQ-MTP-GGUF size only in row label (~12.9 GB) [L66] | M-TI | GGUF | — | 0.104087 ± 0.0016 | — | — | — | ppl:wikitext2 7.0109 | reference not stated | L66 @ main | |
| zerodigest/YMQ-MTP-GGUF size only in row label (~19.0 GB) [L69]KL reference not stated | XL | GGUF | — | 0.011598 ± 0.0005 | — | — | — | ppl:wikitext2 6.8196 | reference not stated | L69 @ main | |
| Zynerji/Ektome-PristinelyUncensored quality figures are for the bf16 abliterated parent, not for any quant [L103] | Ektome (this repo, bf16 safetensors) | other | 54.00 | download | — | — | — | — | MMLU-val 0.818 | other or no reference | L43 @ main |
| Zynerji/Ektome-PristinelyUncensored-GPTQ-MTP MMLU delta -0.015 within SE 0.0195 [L148]receipts measured on the 4-bit weights [L140-143]L231-232 limitation says receipts are from bf16 parent, contradicting L142-143weights 17.7 GiB in prose [L93]; 19.56 GB table size [L38]calibration: 512 wikitext-2 samples, g128 [L225] | this 4-bit build (GPTQ-MTP) | GPTQ | 19.56 | download | — | — | — | — | MMLU-val 0.8025; RefusalCompliance 0.990 | vs BF16 | L38 @ main |
| 0xBakeer/TandemLLM-EXL3-2.5 reference perplexity [L18]publisher comparison build | BF16 reference | EXL3 | — | — | — | — | — | ppl:wikitext-2 perplexity is 6.849. Speed is tokens per | vs BF16 | L18 @ main | |
| 0xBakeer/TandemLLM-EXL3-4.5 [L18]publisher comparison build | BF16 | EXL3 | — | — | — | — | — | ppl:wikitext-2 6.849 | vs BF16 | L18 @ main | |
| 0xBakeer/TandemLLM-EXL3-6.0 BF16 perplexity 6.849 stated in prose [L18]publisher comparison build | BF16 reference | EXL3 | — | — | — | — | — | ppl:wikitext2 perplexity is 6.849. Speed is tokens per | vs BF16 | L18 @ main | |
| 0xBakeer/TandemLLM-NVFP4 double-rounded via FP8 release [L29]publisher comparison build | old full set (from FP8 release) | NVFP4 | — | — | — | — | — | loss_delta:prose +0.0284; loss_delta:code +0.0507 | vs BF16 | L70 @ main | |
| 1bit-MONSTER/Q4_0-H32-GGUF pp512 approximate (~400) [L40]publisher comparison build | Q4_0, exact int8 | GGUF | — | 0.029 | 91.9% | 6.021 | — | — | vs BF16 | L40 @ main | |
| 1bit-MONSTER/Q4_0-H32-GGUF pp512 given as a range 461-488, not mapped [L41]publisher comparison build | Q4_0, W4A4 without rotation | GGUF | — | 0.084 | 87.4% | 6.241 | — | — | vs BF16 | L41 @ main | |
| 6block/GGUF not in this repo [L59]publisher comparison build | master BF16 | GGUF | 53.81 | download | — | — | 6.7856 ± 0.07388 | — | — | vs BF16 | L59 @ main |
| abhishekchohan/AWQ-INT4-FP8KV weights byte-identical to this model; differs only in KV dtype [L63-65]publisher comparison build | Qwen3.8-27B-AWQ-INT4 (bf16 KV) | AWQ | — | — | — | — | — | GSM8K 95.45; GPQA Diamond 61.62; Minerva Math500 80.0; ARC-Easy 82.37; ARC-Challenge 56.83; HellaSwag 55.85; Winogrande 76.48; TruthfulQA MC1 35.50 | other or no reference | L69 @ main | |
| abhishekchohan/GPTQ-INT4 unquantized base on same harness [L61]publisher comparison build | BF16 base | GPTQ | — | — | — | — | — | GSM8K 95.60; GPQA-Diamond 64.65; RealWorldQA 84.31; MathVision 78.06 | vs BF16 | L67 @ main | |
| airagrp/mlx-mxfp8-L [L78]publisher comparison build | Qwen3.8-27B (bf16 reference) | MLX | 55.57 | download | — | — | 6.935 | — | — | vs BF16 | L78 @ main |
| airagrp/mlx-mxfp8-L [L86]publisher comparison build | mlx-mxfp8-XL | MLX | 38.99 | download | 0.0037 | — | 6.937 | — | — | vs BF16 | L86 @ main |
| airagrp/mlx-mxfp8-L [L84]publisher comparison build | mlx-mxfp8-M | MLX | 31.98 | download | 0.0068 | — | 6.893 | — | — | vs BF16 | L84 @ main |
| airagrp/mlx-mxfp8-L [L83]publisher comparison build | mlx-nvfp4-XL | MLX | 30.97 | download | 0.0382 | — | 7.024 | — | — | vs BF16 | L83 @ main |
| airagrp/mlx-mxfp8-L [L82]publisher comparison build | mlx-mxfp8-S | MLX | 28.58 | download | 0.0278 | — | 6.922 | — | — | vs BF16 | L82 @ main |
| airagrp/mlx-mxfp8-L [L81]publisher comparison build | mlx-nvfp4-L | MLX | 23.96 | download | 0.0407 | — | 6.987 | — | — | vs BF16 | L81 @ main |
| airagrp/mlx-mxfp8-L [L80]publisher comparison build | mlx-nvfp4-M | MLX | 23.16 | download | 0.0490 | — | 6.983 | — | — | vs BF16 | L80 @ main |
| airagrp/mlx-mxfp8-L [L79]publisher comparison build | mlx-nvfp4-S | MLX | 20.56 | download | 0.0573 | — | 7.024 | — | — | vs BF16 | L79 @ main |
| airagrp/mlx-mxfp8-M publisher comparison build | Qwen3.8-27B (bf16 reference) | MLX | 55.57 | download | — | — | 6.935 | — | — | vs BF16 | L78 @ main |
| airagrp/mlx-mxfp8-M publisher comparison build | Qwen3.8-27B-mlx-mxfp8-XL | MLX | 38.99 | download | 0.0037 | — | 6.937 | — | — | vs BF16 | L86 @ main |
| airagrp/mlx-mxfp8-M publisher comparison build | Qwen3.8-27B-mlx-mxfp8-L | MLX | 37.36 | download | 0.0046 | — | 6.937 | — | — | vs BF16 | L85 @ main |
| airagrp/mlx-mxfp8-M publisher comparison build | Qwen3.8-27B-mlx-nvfp4-XL | MLX | 30.97 | download | 0.0382 | — | 7.024 | — | — | vs BF16 | L83 @ main |
| airagrp/mlx-mxfp8-M publisher comparison build | Qwen3.8-27B-mlx-mxfp8-S | MLX | 28.58 | download | 0.0278 | — | 6.922 | — | — | vs BF16 | L82 @ main |
| airagrp/mlx-mxfp8-M publisher comparison build | Qwen3.8-27B-mlx-nvfp4-L | MLX | 23.96 | download | 0.0407 | — | 6.987 | — | — | vs BF16 | L81 @ main |
| airagrp/mlx-mxfp8-M publisher comparison build | Qwen3.8-27B-mlx-nvfp4-M | MLX | 23.16 | download | 0.0490 | — | 6.983 | — | — | vs BF16 | L80 @ main |
| airagrp/mlx-mxfp8-M publisher comparison build | Qwen3.8-27B-mlx-nvfp4-S | MLX | 20.56 | download | 0.0573 | — | 7.024 | — | — | vs BF16 | L79 @ main |
| airagrp/mlx-mxfp8-S reference [L79]publisher comparison build | Qwen3.8-27B (bf16 reference) | MLX | 55.57 | download | — | — | — | — | ppl:wikitext-2 6.935 | vs BF16 | L79 @ main |
| airagrp/mlx-mxfp8-S publisher comparison build | Qwen3.8-27B-mlx-mxfp8-XL | MLX | 38.99 | download | 0.0037 | — | — | — | ppl:wikitext-2 6.937 | vs BF16 | L87 @ main |
| airagrp/mlx-mxfp8-S publisher comparison build | Qwen3.8-27B-mlx-mxfp8-L | MLX | 37.36 | download | 0.0046 | — | — | — | ppl:wikitext-2 6.937 | vs BF16 | L86 @ main |
| airagrp/mlx-mxfp8-S publisher comparison build | Qwen3.8-27B-mlx-mxfp8-M | MLX | 31.98 | download | 0.0068 | — | — | — | ppl:wikitext-2 6.893 | vs BF16 | L85 @ main |
| airagrp/mlx-mxfp8-S publisher comparison build | Qwen3.8-27B-mlx-nvfp4-XL | MLX | 30.97 | download | 0.0382 | — | — | — | ppl:wikitext-2 7.024 | vs BF16 | L84 @ main |
| airagrp/mlx-mxfp8-S publisher comparison build | Qwen3.8-27B-mlx-nvfp4-L | MLX | 23.96 | download | 0.0407 | — | — | — | ppl:wikitext-2 6.987 | vs BF16 | L82 @ main |
| airagrp/mlx-mxfp8-S publisher comparison build | Qwen3.8-27B-mlx-nvfp4-M | MLX | 23.16 | download | 0.0490 | — | — | — | ppl:wikitext-2 6.983 | vs BF16 | L81 @ main |
| airagrp/mlx-mxfp8-S publisher comparison build | Qwen3.8-27B-mlx-nvfp4-S | MLX | 20.56 | download | 0.0573 | — | — | — | ppl:wikitext-2 7.024 | vs BF16 | L80 @ main |
| airagrp/mlx-mxfp8-XL [L78]publisher comparison build | Qwen3.8-27B (bf16 reference) | MLX | 55.57 | download | — | — | 6.935 | — | — | vs BF16 | L78 @ main |
| airagrp/mlx-mxfp8-XL [L85]publisher comparison build | Qwen3.8-27B-mlx-mxfp8-L | MLX | 37.36 | download | 0.0046 | — | 6.937 | — | — | vs BF16 | L85 @ main |
| airagrp/mlx-mxfp8-XL [L84]publisher comparison build | Qwen3.8-27B-mlx-mxfp8-M | MLX | 31.98 | download | 0.0068 | — | 6.893 | — | — | vs BF16 | L84 @ main |
| airagrp/mlx-mxfp8-XL [L83]publisher comparison build | Qwen3.8-27B-mlx-nvfp4-XL | MLX | 30.97 | download | 0.0382 | — | 7.024 | — | — | vs BF16 | L83 @ main |
| airagrp/mlx-mxfp8-XL [L82]publisher comparison build | Qwen3.8-27B-mlx-mxfp8-S | MLX | 28.58 | download | 0.0278 | — | 6.922 | — | — | vs BF16 | L82 @ main |
| airagrp/mlx-mxfp8-XL [L81]publisher comparison build | Qwen3.8-27B-mlx-nvfp4-L | MLX | 23.96 | download | 0.0407 | — | 6.987 | — | — | vs BF16 | L81 @ main |
| airagrp/mlx-mxfp8-XL [L80]publisher comparison build | Qwen3.8-27B-mlx-nvfp4-M | MLX | 23.16 | download | 0.0490 | — | 6.983 | — | — | vs BF16 | L80 @ main |
| airagrp/mlx-mxfp8-XL [L79]publisher comparison build | Qwen3.8-27B-mlx-nvfp4-S | MLX | 20.56 | download | 0.0573 | — | 7.024 | — | — | vs BF16 | L79 @ main |
| airagrp/mlx-nvfp4-L reference [L79]publisher comparison build | Qwen3.8-27B (bf16 reference) | MLX | 55.57 | unknown | — | — | 6.935 | — | — | vs BF16 | L79 @ main |
| airagrp/mlx-nvfp4-L [L87]publisher comparison build | Qwen3.8-27B-mlx-mxfp8-XL | MLX | 38.99 | unknown | 0.0037 | — | 6.937 | — | — | vs BF16 | L87 @ main |
| airagrp/mlx-nvfp4-L [L86]publisher comparison build | Qwen3.8-27B-mlx-mxfp8-L | MLX | 37.36 | unknown | 0.0046 | — | 6.937 | — | — | vs BF16 | L86 @ main |
| airagrp/mlx-nvfp4-L [L85]publisher comparison build | Qwen3.8-27B-mlx-mxfp8-M | MLX | 31.98 | unknown | 0.0068 | — | 6.893 | — | — | vs BF16 | L85 @ main |
| airagrp/mlx-nvfp4-L [L84]publisher comparison build | Qwen3.8-27B-mlx-nvfp4-XL | MLX | 30.97 | unknown | 0.0382 | — | 7.024 | — | — | vs BF16 | L84 @ main |
| airagrp/mlx-nvfp4-L [L83]publisher comparison build | Qwen3.8-27B-mlx-mxfp8-S | MLX | 28.58 | unknown | 0.0278 | — | 6.922 | — | — | vs BF16 | L83 @ main |
| airagrp/mlx-nvfp4-L [L81]publisher comparison build | Qwen3.8-27B-mlx-nvfp4-M | MLX | 23.16 | unknown | 0.0490 | — | 6.983 | — | — | vs BF16 | L81 @ main |
| airagrp/mlx-nvfp4-L [L80]publisher comparison build | Qwen3.8-27B-mlx-nvfp4-S | MLX | 20.56 | unknown | 0.0573 | — | 7.024 | — | — | vs BF16 | L80 @ main |
| airagrp/mlx-nvfp4-M publisher comparison build | Qwen3.8-27B bf16 reference | MLX | 55.57 | download | — | — | — | — | ppl:wikitext2 6.935 | vs BF16 | L79 @ main |
| airagrp/mlx-nvfp4-M publisher comparison build | mlx-mxfp8-XL | MLX | 38.99 | download | — | — | — | — | ppl:wikitext2 6.937; kld:wikitext2 0.0037 | vs BF16 | L87 @ main |
| airagrp/mlx-nvfp4-M publisher comparison build | mlx-mxfp8-L | MLX | 37.36 | download | — | — | — | — | ppl:wikitext2 6.937; kld:wikitext2 0.0046 | vs BF16 | L86 @ main |
| airagrp/mlx-nvfp4-M publisher comparison build | mlx-mxfp8-M | MLX | 31.98 | download | — | — | — | — | ppl:wikitext2 6.893; kld:wikitext2 0.0068 | vs BF16 | L85 @ main |
| airagrp/mlx-nvfp4-M publisher comparison build | mlx-nvfp4-XL | MLX | 30.97 | download | — | — | — | — | ppl:wikitext2 7.024; kld:wikitext2 0.0382 | vs BF16 | L84 @ main |
| airagrp/mlx-nvfp4-M publisher comparison build | mlx-mxfp8-S | MLX | 28.58 | download | — | — | — | — | ppl:wikitext2 6.922; kld:wikitext2 0.0278 | vs BF16 | L83 @ main |
| airagrp/mlx-nvfp4-M publisher comparison build | mlx-nvfp4-L | MLX | 23.96 | download | — | — | — | — | ppl:wikitext2 6.987; kld:wikitext2 0.0407 | vs BF16 | L82 @ main |
| airagrp/mlx-nvfp4-M publisher comparison build | mlx-nvfp4-S | MLX | 20.56 | download | — | — | — | — | ppl:wikitext2 7.024; kld:wikitext2 0.0573 | vs BF16 | L80 @ main |
| airagrp/mlx-nvfp4-S publisher comparison build | Qwen3.8-27B (bf16 reference) | MLX | 55.57 | download | — | — | 6.935 | — | — | vs BF16 | L78 @ main |
| airagrp/mlx-nvfp4-S publisher comparison build | Qwen3.8-27B-mlx-mxfp8-XL | MLX | 38.99 | download | 0.0037 | — | 6.937 | — | — | vs BF16 | L86 @ main |
| airagrp/mlx-nvfp4-S publisher comparison build | Qwen3.8-27B-mlx-mxfp8-L | MLX | 37.36 | download | 0.0046 | — | 6.937 | — | — | vs BF16 | L85 @ main |
| airagrp/mlx-nvfp4-S publisher comparison build | Qwen3.8-27B-mlx-mxfp8-M | MLX | 31.98 | download | 0.0068 | — | 6.893 | — | — | vs BF16 | L84 @ main |
| airagrp/mlx-nvfp4-S publisher comparison build | Qwen3.8-27B-mlx-nvfp4-XL | MLX | 30.97 | download | 0.0382 | — | 7.024 | — | — | vs BF16 | L83 @ main |
| airagrp/mlx-nvfp4-S publisher comparison build | Qwen3.8-27B-mlx-mxfp8-S | MLX | 28.58 | download | 0.0278 | — | 6.922 | — | — | vs BF16 | L82 @ main |
| airagrp/mlx-nvfp4-S publisher comparison build | Qwen3.8-27B-mlx-nvfp4-L | MLX | 23.96 | download | 0.0407 | — | 6.987 | — | — | vs BF16 | L81 @ main |
| airagrp/mlx-nvfp4-S publisher comparison build | Qwen3.8-27B-mlx-nvfp4-M | MLX | 23.16 | download | 0.0490 | — | 6.983 | — | — | vs BF16 | L80 @ main |
| airagrp/mlx-nvfp4-XL publisher comparison build | bf16 reference | MLX | 55.57 | download | — | — | — | — | ppl:wikitext2 6.935 | vs BF16 | L78 @ main |
| airagrp/mlx-nvfp4-XL publisher comparison build | mlx-mxfp8-XL | MLX | 38.99 | download | 0.0037 | — | — | — | ppl:wikitext2 6.937 | vs BF16 | L86 @ main |
| airagrp/mlx-nvfp4-XL publisher comparison build | mlx-mxfp8-L | MLX | 37.36 | download | 0.0046 | — | — | — | ppl:wikitext2 6.937 | vs BF16 | L85 @ main |
| airagrp/mlx-nvfp4-XL publisher comparison build | mlx-mxfp8-M | MLX | 31.98 | download | 0.0068 | — | — | — | ppl:wikitext2 6.893 | vs BF16 | L84 @ main |
| airagrp/mlx-nvfp4-XL publisher comparison build | mlx-mxfp8-S | MLX | 28.58 | download | 0.0278 | — | — | — | ppl:wikitext2 6.922 | vs BF16 | L82 @ main |
| airagrp/mlx-nvfp4-XL publisher comparison build | mlx-nvfp4-L | MLX | 23.96 | download | 0.0407 | — | — | — | ppl:wikitext2 6.987 | vs BF16 | L81 @ main |
| airagrp/mlx-nvfp4-XL publisher comparison build | mlx-nvfp4-M | MLX | 23.16 | download | 0.0490 | — | — | — | ppl:wikitext2 6.983 | vs BF16 | L80 @ main |
| airagrp/mlx-nvfp4-XL publisher comparison build | mlx-nvfp4-S | MLX | 20.56 | download | 0.0573 | — | — | — | ppl:wikitext2 7.024 | vs BF16 | L79 @ main |
| akopytko/NVFP4-GGUF author-made plain Q4_0 [L62,L69]publisher comparison build | Q4_0 w/o imatrix | GGUF | 15.71 | download | 0.037877 +/- 0.000448 | 91.461 +/- 0.124% | — | — | — | vs BF16 | L62 @ main |
| Alogotron/MixQ4A6E8-GGUF [L100]publisher comparison build | Q6_K | GGUF | 22.08 | download | 0.002099 | — | 6.6553 | — | — | vs BF16 | L100 @ main |
| Alogotron/MixQ4A6E8-GGUF sibling release [L99]publisher comparison build | MixQ5A6E8-v0 (sibling) | GGUF | 20.80 | download | 0.003083 | — | 6.6674 | — | — | vs BF16 | L99 @ main |
| Alogotron/MixQ4A6E8-GGUF [L98]publisher comparison build | Q5_K_M | GGUF | 19.23 | download | 0.003949 | — | 6.6715 | — | — | vs BF16 | L98 @ main |
| Alogotron/MixQ5A6E8-GGUF publisher comparison build | Q6_K | GGUF | 22.08 | download | 0.002099 | — | 6.6553 | — | — | vs BF16 | L105 @ main |
| Alogotron/MixQ5A6E8-GGUF publisher comparison build | Q5_K_M | GGUF | 19.23 | download | 0.003949 | — | 6.6715 | — | — | vs BF16 | L103 @ main |
| Alogotron/MixQ5A6E8-GGUF publisher comparison build | MixQ4A6E8-v0 (sibling candidate) | GGUF | 19.02 | download | 0.007835 | — | 6.6830 | — | — | vs BF16 | L102 @ main |
| Alogotron/MixQ5A6E8-GGUF publisher comparison build | Q4_K_M | GGUF | 17.77 | download | 0.009584 | — | 6.6921 | — | — | vs BF16 | L101 @ main |
| amd/Quark-AWQ-INT4-W4A16 reference measured by the card author [L36]publisher comparison build | BF16 base | AWQ | — | — | — | 8.4364 ppl:wikitext | — | BFCL_single_turn 24.38% | vs BF16 | L29 @ main | |
| anemll/anemll-forge- [L60-67]publisher comparison build | ANEMLL previous packages / release 1 (ANE M6) | other | — | 0.1838 | 86.0% | 2.414 | — | kld_99 2.28; ppl:verify_path 6.645; ppl:8k 8.096; ppl:64k 5.068 | vs BF16 | L60 @ main | |
| Ar3ntiz/qwen3_27B_side8 4.22 bpp from row label, no candidate id [L56]publisher comparison build | SUC1 base (4.22 bpp) | other | 14.18 | download | 0.031 | — | — | — | — | vs BF16 | L56 @ main |
| armin1/NVFP4-FP8-Mixed-LH recipe ablation step, publisher's own build [L69]publisher comparison build | RadixArk recipe reproduced bit for bit (max calibration) | NVFP4 | — | 0.0242 | — | — | — | — | vs BF16 | L69 @ main | |
| armin1/NVFP4-FP8-Mixed-LH recipe ablation step [L70]publisher comparison build | NVIDIA Local-Hessian recipe + BF16 lm_head | NVFP4 | — | 0.0209 | — | — | — | — | vs BF16 | L70 @ main | |
| armin1/NVFP4-FP8-Mixed-LH recipe ablation step [L71]publisher comparison build | + last 8 MLPs in FP8 | NVFP4 | — | 0.0171 | — | — | — | — | vs BF16 | L71 @ main | |
| armin1/NVFP4-FP8-Mixed-LH-Heretic Publisher's non-abliterated sibling, same recipe [L110, L123]publisher comparison build | NVFP4-FP8-Mixed-LH (not abliterated) | NVFP4 | — | 0.0178 | 95.98% | — | — | kld_99 0.206; GSM8K 96.06% | vs BF16 | L110 @ main | |
| AtomicChat/GGUF [L239]publisher comparison build | layout: 16 lifted, plus a richer output head | GGUF | 18.80 | download | 0.00800 | — | — | — | — | vs BF16 | L239 @ main |
| AtomicChat/GGUF [L240]publisher comparison build | layout: 24 layers lifted | GGUF | 18.60 | download | 0.00743 | — | — | — | — | vs BF16 | L240 @ main |
| AtomicChat/GGUF [L237]publisher comparison build | layout: 32 layers lifted instead | GGUF | 18.40 | download | 0.00826 | — | — | — | — | vs BF16 | L237 @ main |
| AtomicChat/GGUF [L238]publisher comparison build | layout: 16 lifted, plus the attention gate | GGUF | 18.40 | download | 0.00821 | — | — | — | — | vs BF16 | L238 @ main |
| AtomicChat/GGUF [L235]publisher comparison build | layout: more bits on attention | GGUF | 18.20 | download | 0.01010 | — | — | — | — | vs BF16 | L235 @ main |
| AtomicChat/GGUF [L234]publisher comparison build | layout: more bits on ffn_down everywhere | GGUF | 17.80 | download | 0.01189 | — | — | — | — | vs BF16 | L234 @ main |
| AtomicChat/GGUF [L236]publisher comparison build | layout: 16 layers lifted, first and last | GGUF | 17.80 | download | 0.00981 | — | — | — | — | vs BF16 | L236 @ main |
| AtomicChat/GGUF [L233]publisher comparison build | layout: 4 layers lifted | GGUF | 17.10 | download | 0.01449 | — | — | — | — | vs BF16 | L233 @ main |
| AtomicChat/GGUF layout experiment [L232]publisher comparison build | layout: every layer treated the same | GGUF | 16.80 | download | 0.01580 | — | — | — | — | vs BF16 | L232 @ main |
| Avesed/INT4-W4A16 GSM8K 96.8% (242/250) same protocol [L46-47]publisher comparison build | INT8-W8A8 | INT4 | — | — | — | — | — | GSM8K 96.8 % | other or no reference | L46 @ main | |
| Avesed/INT8-W8A8 GSM8K 96.4% (241/250) identical protocol [L35]publisher comparison build | INT4-W4A16 | INT8 | — | — | — | — | — | GSM8K 96.4 % | other or no reference | L35 @ main | |
| avlp12/Alis-MLX-4bit [L342]publisher comparison build | bf16 reference | MLX | 51.80 | download | — | — | — | — | ppl:en 5.7734; ppl:ko 6.0954; ppl:code 1.6813 | vs BF16 | L33 @ main |
| avlp12/Alis-MLX-4bit PPL on pre-graded era weights per dagger [L343]publisher comparison build | 8-bit | MLX | 27.90 | download | — | — | — | — | ppl:en 5.7760; ppl:ko 6.0987; ppl:code 1.6815 | vs BF16 | L34 @ main |
| avlp12/Alis-MLX-4bit [L370]publisher comparison build | 4-bit AWQ, pre-graded (pre-graded branch) | MLX | 15.17 | download | 0.0649 | 91.3% | — | — | — | vs BF16 | L370 @ main |
| avlp12/Alis-MLX-6bit bf16 row carried over from earlier batch [L33]publisher comparison build | bf16 reference | MLX | 51.80 | download | — | — | — | — | ppl:en 5.7734; ppl:ko 6.0954; ppl:code 1.6813 | vs BF16 | L28 @ main |
| avlp12/Alis-MLX-6bit PPL measured on pre-graded weights [L308]exact KL 0.00105 in prose [L316]publisher comparison build | 8-bit | MLX | 27.90 | download | 0.00105 | — | — | — | ppl:en 5.7760; ppl:ko 6.0987; ppl:code 1.6815 | vs BF16 | L29 @ main |
| avlp12/Alis-MLX-6bit PPL measured on pre-graded weights [L308]exact KL 0.0592 in prose [L316]duplicate table at L291-296 is misparsed; L28-31 table usedpublisher comparison build | 4-bit AWQ | MLX | 15.40 | download | 0.0592 | — | — | — | ppl:en 5.8450; ppl:ko 6.2609; ppl:code 1.8105 | vs BF16 | L31 @ main |
| avlp12/Alis-MLX-8bit reference row [L294]publisher comparison build | bf16 reference | MLX | 51.80 | download | — | — | — | — | ppl:en 5.7734; ppl:ko 6.0954; ppl:code 1.6813 | vs BF16 | L28 @ main |
| avlp12/Alis-MLX-8bit PPL on pre-graded weights [L303]publisher comparison build | 6-bit | MLX | 21.90 | download | — | — | — | — | ppl:en 5.7924; ppl:ko 6.1018; ppl:code 1.6854 | vs BF16 | L30 @ main |
| avlp12/Alis-MLX-8bit quality-table row for 4-bit is mis-rendered; values taken from at-a-glance table [L31]publisher comparison build | 4-bit AWQ | MLX | 15.40 | download | — | — | — | — | ppl:en 5.8450; ppl:ko 6.2609; ppl:code 1.8105; ppl_delta_pct:en +1.2; ppl_delta_pct:ko +2.7; ppl_delta_pct:code +7.7 | vs BF16 | L31 @ main |
| AxionML/NVFP4 (scores reported by RadixArk, not named repo of their own) quoted by AxionML, not measuredScores reported by RadixArk on 4x B300/GB300, SGLang TP4, temperature 1.0 [L47]GSM8K thinking mode, text-only; Terminal-Bench is an 84-task subset pass@1 [L44-45]No BF16 comparison scores on card | NVFP4 (mixed NVFP4/FP8) | NVFP4 | — | — | — | — | — | GSM8K 97.27; Terminal-Bench 2.1 73.81 | reference not stated | L44 @ main | |
| baa-ai/RAM-13GB-GGUF [L84]publisher comparison build | v1 (llama.cpp IQ3_M) | GGUF | 12.58 | download | — | — | — | — | ppl:wikitext2 6.34 | reference not stated | L84 @ main |
| baa-ai/RAM-24GB-MLX unquantized baseline measured by the card author [L47,L69]publisher comparison build | Qwen3.8-27B BF16 (full precision) | MLX | 55.60 | download | — | — | — | — | canary_invented_x 0.000; canary_recall 0.999; canary_branch 0.986 | vs BF16 | L47 @ main |
| baa-ai/RAM-24GB-MLX quoted by baa-ai, not measuredMMLU 90.0% matching BF16, quoted from the MLX sibling card, different artifact [L54] | Qwen3.8-27B-RAM-24GB-MLX (sibling) | GGUF | — | — | — | — | — | MMLU 90.0% | vs BF16 | L54 @ main | |
| bartowski/GGUF previous release, standard layout [L196-200]publisher comparison build | Q6_K_L (previous release) | GGUF | 24.08 | download | 0.0027 ± 0.0001 | 97.64 ± 0.10 % | — | — | — | vs BF16 | L200 @ main |
| bartowski/GGUF previous release [L201]publisher comparison build | Q6_K (previous release) | GGUF | 23.46 | download | 0.0031 ± 0.0001 | 97.49 ± 0.10 % | — | — | — | vs BF16 | L201 @ main |
| bartowski/GGUF previous release only, no current counterpart [L203]publisher comparison build | Q5_K_L (previous release) | GGUF | 21.54 | download | 0.0050 ± 0.0002 | 96.92 ± 0.11 % | — | — | — | vs BF16 | L203 @ main |
| bartowski/GGUF previous release [L204]publisher comparison build | Q5_K_M (previous release) | GGUF | 20.75 | download | 0.0054 ± 0.0001 | 96.71 ± 0.11 % | — | — | — | vs BF16 | L204 @ main |
| bartowski/GGUF previous release [L205]publisher comparison build | Q5_K_S (previous release) | GGUF | 19.68 | download | 0.0065 ± 0.0001 | 96.25 ± 0.12 % | — | — | — | vs BF16 | L205 @ main |
| bartowski/GGUF previous release [L206]publisher comparison build | Q4_K_L (previous release) | GGUF | 18.72 | download | 0.0124 ± 0.0003 | 95.00 ± 0.14 % | — | — | — | vs BF16 | L206 @ main |
| bartowski/GGUF previous release [L207]publisher comparison build | Q4_K_M (previous release) | GGUF | 17.77 | download | 0.0129 ± 0.0002 | 95.02 ± 0.14 % | — | — | — | vs BF16 | L207 @ main |
| bartowski/GGUF previous release [L210]publisher comparison build | Q4_K_S (previous release) | GGUF | 16.71 | download | 0.0165 ± 0.0003 | 94.18 ± 0.15 % | — | — | — | vs BF16 | L210 @ main |
| bartowski/GGUF previous release only [L209]publisher comparison build | Q3_K_XL (previous release) | GGUF | 16.39 | download | 0.0426 ± 0.0007 | 91.01 ± 0.18 % | — | — | — | vs BF16 | L209 @ main |
| bartowski/GGUF previous release [L208]publisher comparison build | IQ4_NL (previous release) | GGUF | 16.33 | download | 0.0178 ± 0.0003 | 94.22 ± 0.15 % | — | — | — | vs BF16 | L208 @ main |
| bartowski/GGUF previous release [L211]publisher comparison build | IQ4_XS (previous release) | GGUF | 15.57 | download | 0.0196 ± 0.0005 | 94.03 ± 0.15 % | — | — | — | vs BF16 | L211 @ main |
| bartowski/GGUF previous release [L213]publisher comparison build | Q3_K_L (previous release) | GGUF | 15.28 | download | 0.0437 ± 0.0007 | 90.82 ± 0.18 % | — | — | — | vs BF16 | L213 @ main |
| bartowski/GGUF previous release [L214]publisher comparison build | Q3_K_M (previous release) | GGUF | 14.61 | download | 0.0466 ± 0.0008 | 90.57 ± 0.18 % | — | — | — | vs BF16 | L214 @ main |
| bartowski/GGUF previous release [L212]publisher comparison build | IQ3_M (previous release) | GGUF | 13.90 | download | 0.0566 ± 0.0008 | 89.66 ± 0.19 % | — | — | — | vs BF16 | L212 @ main |
| bartowski/GGUF previous release [L217]publisher comparison build | Q3_K_S (previous release) | GGUF | 13.72 | download | 0.0686 ± 0.0010 | 88.17 ± 0.20 % | — | — | — | vs BF16 | L217 @ main |
| bartowski/GGUF previous release [L216]publisher comparison build | IQ3_XS (previous release) | GGUF | 13.33 | download | 0.0624 ± 0.0009 | 89.18 ± 0.20 % | — | — | — | vs BF16 | L216 @ main |
| bartowski/GGUF previous release only [L215]publisher comparison build | Q2_K_L (previous release) | GGUF | 13.08 | download | 0.1322 ± 0.0018 | 84.37 ± 0.23 % | — | — | — | vs BF16 | L215 @ main |
| bartowski/GGUF previous release [L218]publisher comparison build | IQ3_XXS (previous release) | GGUF | 12.63 | download | 0.0850 ± 0.0012 | 87.34 ± 0.21 % | — | — | — | vs BF16 | L218 @ main |
| bartowski/GGUF previous release [L219]publisher comparison build | Q2_K (previous release) | GGUF | 11.84 | download | 0.1357 ± 0.0019 | 83.98 ± 0.23 % | — | — | — | vs BF16 | L219 @ main |
| bartowski/GGUF previous release [L220]publisher comparison build | IQ2_M (previous release) | GGUF | 10.87 | download | 0.1551 ± 0.0021 | 83.06 ± 0.23 % | — | — | — | vs BF16 | L220 @ main |
| bartowski/GGUF previous release [L221]publisher comparison build | IQ2_S (previous release) | GGUF | 10.30 | download | 0.2166 ± 0.0026 | 79.78 ± 0.25 % | — | — | — | vs BF16 | L221 @ main |
| bartowski/GGUF previous release [L222]publisher comparison build | IQ2_XS (previous release) | GGUF | 10 | download | 0.2293 ± 0.0028 | 79.53 ± 0.25 % | — | — | — | vs BF16 | L222 @ main |
| bartowski/GGUF previous release [L223]publisher comparison build | IQ2_XXS (previous release) | GGUF | 9.4 | download | 0.3037 ± 0.0034 | 76.29 ± 0.27 % | — | — | — | vs BF16 | L223 @ main |
| bg-digitalservices/ROCmFP4-STRIX-GGUF [L47]publisher comparison build | stock STRIX, no head fix | GGUF | 15.01 | download | — | — | — | — | kld:EN 0.046211; kld:DE 0.042837; top1:EN 88.91% | other or no reference | L47 @ main |
| bg-digitalservices/ROCmFP4-STRIX-GGUF [L48]publisher comparison build | stock STRIX, no imatrix | GGUF | 15.01 | download | — | — | — | — | kld:EN 0.052341; kld:DE 0.048451; top1:EN 88.58% | other or no reference | L48 @ main |
| btbtyler09/GPTQ-4bit reference row [L87]publisher comparison build | BF16 (original) | GPTQ | 56.00 | download | — | — | 6.4457 | — | — | vs BF16 | L76 @ main |
| btbtyler09/GPTQ-8bit Publisher's sibling btbtyler09/Qwen3.8-27B-GPTQ-4bit [L21, L83]publisher comparison build | GPTQ 4-bit FOEM | GPTQ | 21.00 | download | — | — | — | +0.81% | ppl:wikitext-2 6.4982 | vs BF16 | L71 @ main |
| Bucoid/Heretic-Ara-16GB-VRAM-IQ4-XS-MTP-GGUF reference row; KLD/top-1 are self-comparison [L42]publisher comparison build | Heretic-Ara BF16 (base) | GGUF | 53.79 | download | — | — | 7.008212 ± 0.045362 | — | — | vs BF16 | L31 @ main |
| cafonez/ROCmI4-MTP-GGUF reference for ROCmI4 comparison [L207]publisher comparison build | local Q4_K_M reference | GGUF | — | — | — | 5.9455 | — | — | other or no reference | L207 @ main | |
| Chungulus/MLX-3bit-Group32 [L125]publisher comparison build | BF16 reference | MLX | — | — | — | perplexity": 9.84727010437735, | — | — | vs BF16 | L125 @ main | |
| Chungulus/MLX-5bit-Group32 reference_perplexity [L125]publisher comparison build | BF16 reference | MLX | — | — | — | perplexity": 9.84727010437735, | — | — | vs BF16 | L125 @ main | |
| Chungulus/MLX-8bit-Group32 reference perplexity [L125]publisher comparison build | BF16 reference | MLX | — | — | — | perplexity": 9.84727010437735, | — | — | vs BF16 | L125 @ main | |
| cloudnathan5/NVFP4a4-GPTQ unquantized reference [L97]publisher comparison build | Qwen/Qwen3.8-27B (BF16) | NVFP4 | — | — | — | 6.5574 | — | — | vs BF16 | L97 @ main | |
| curvedinf/GPTQ-INT8-W8A8-GS128-PTQR KLD 0.0110, greedy agreement 38/52 [L61]publisher comparison build | original GPTQ checkpoint | GPTQ | — | 0.0110 | — | — | — | — | vs BF16 | L61 @ main | |
| cyankiwi/AWQ-INT4 measured by malaiwahvLLM, shard-0 [L462] | cyankiwi/Qwen3.8-27B-AWQ-INT4 | AWQ | 21.01 | download | 0.022818 | 93.94 % | — | — | — | vs BF16 | L462 @ main |
| cyankiwi/AWQ-INT4 measured by malaiwahshard-0 [L471] | cyankiwi/Qwen3.8-27B-AWQ-INT4 | AWQ | 21.01 | download | 0.022818 | 93.94 % | — | — | — | vs BF16 | L471 @ main |
| cyankiwi/AWQ-INT4 measured by malaiwah[L502] | cyankiwi/Qwen3.8-27B-AWQ-INT4 shard0 | AWQ | 21.01 | download | 0.022818 | 93.94 % | — | — | — | other or no reference | L502 @ main |
| cyankiwi/AWQ-INT4 measured by TelperionAI | cyankiwi/Qwen3.8-27B-AWQ-INT4 | AWQ | 21.00 | download | — | 94.35% | — | — | — | vs BF16 | L56 @ main |
| cyankiwi/AWQ-INT4 measured by TelperionAI[L61] | cyankiwi/Qwen3.8-27B-AWQ-INT4 | AWQ | 21.00 | download | — | 94.35% | — | — | near-tie disagree 33.74%; moderate disagree 5.59%; confident disagree 1.35%; certain disagree 0.12%; divmed 29 | vs BF16 | L61 @ main |
| CySpiegel/Int4-AutoRound baseline served on same vLLM build [L39]publisher comparison build | FP8 (online per-tensor W8A16) | AutoRound | — | — | — | — | — | GSM8K 5-shot 38.4 | other or no reference | L44 @ main | |
| dbirks/W4A16-AutoRound Publisher's BF16 baseline [L81-85]publisher comparison build | BF16 base | AutoRound | — | — | — | — | — | GSM8K 0.911 ±0.015; HumanEval 0.939 ±0.037; MMLU-Pro 0.819 ±0.020 | vs BF16 | L83 @ main | |
| dbirks/W4A16-AutoRound quoted by fasolack, not measuredColumn header says BF16 but prose attributes 0.917 / 0.826 to the reference W4A16 build vs 0.911 / 0.819 BF16, so the column is ambiguous [L37-39,L276-279]Reference decoding settings never disclosed [L257-258] | Reference (dbirks) | INT4 | — | — | — | — | — | GSM8K 91.7%; MMLU-Pro 82.6%; HumanEval 0.957 | other or no reference | L276 @ main | |
| ddalcu/MLX-Serve-6bit Sibling build by same publisher [L45]; standard error 0.4-0.9 points [L50]publisher comparison build | 8-bit | MLX | 31.20 | download | 0.058 | 96.4% | — | — | — | vs BF16 | L45 @ main |
| ddalcu/MLX-Serve-6bit Sibling build by same publisher [L47]publisher comparison build | 4-bit | MLX | 18.20 | download | 0.496 | 83.6% | — | — | — | vs BF16 | L47 @ main |
| ddalcu/MLX-Serve-6bit Sibling imatrix mixed-width build [L48]publisher comparison build | iQ-MLX 3.8bpw | MLX | 13.00 | download | 0.611 | 79.6% | — | — | — | vs BF16 | L48 @ main |
| ddalcu/MLX-Serve-iQ-MLX-3.8bpw publisher's own 8-bit build [L45]publisher comparison build | 8-bit | MLX | 30.70 | resident | 0.014 | 95.5% | — | — | — | vs BF16 | L45 @ main |
| ddalcu/MLX-Serve-iQ-MLX-3.8bpw publisher's own 6-bit build [L46]publisher comparison build | 6-bit | MLX | 22.70 | resident | 0.056 | 95.5% | — | — | — | vs BF16 | L46 @ main |
| ddalcu/MLX-Serve-iQ-MLX-3.8bpw publisher's own 4-bit build [L47]publisher comparison build | 4-bit | MLX | 17.70 | resident | 0.322 | 83.6% | — | — | — | vs BF16 | L47 @ main |
| dealignai/JANG_2D-CRACK CRACK MMLU cell not a numeric candidate, delta +0.0pp [L121]publisher comparison build | MXFP8 CRACK | other | 27.00 | download | — | — | — | — | MMLU_base 86.67% | other or no reference | L121 @ main |
| dealignai/JANG_2D-CRACK [L120]publisher comparison build | JANG_6D CRACK | other | 24.00 | download | — | — | — | — | MMLU_base 88.77% | other or no reference | L120 @ main |
| dealignai/JANG_2D-CRACK [L119]publisher comparison build | JANG_4D CRACK | other | 17.00 | download | — | — | — | — | MMLU_base 88.77% | other or no reference | L119 @ main |
| dealignai/JANG_4D-CRACK [L120]publisher comparison build | 6D | other | 24.00 | download | — | — | — | — | MMLU 89.12% | other or no reference | L120 @ main |
| dealignai/JANG_4D-CRACK MMLU base 80.0% [L118]publisher comparison build | 2D | other | 11.00 | download | — | — | — | — | MMLU 76.84% | other or no reference | L118 @ main |
| dealignai/JANG_4D-CRACK [L119]publisher comparison build | Base (unmodified, MMLU on 4D harness row) | other | — | — | — | — | — | MMLU 88.77% | other or no reference | L119 @ main | |
| dealignai/JANG_6D-CRACK MMLU CRACK cell absent in table; delta +0.0pp [L121]publisher comparison build | MXFP8 | other | 27.00 | download | — | — | — | — | MMLU_base 86.67% | other or no reference | L121 @ main |
| dealignai/JANG_6D-CRACK publisher comparison build | 4D | other | 17.00 | download | — | — | — | — | MMLU 87.72%; MMLU_base 88.77% | other or no reference | L119 @ main |
| dealignai/JANG_6D-CRACK publisher comparison build | 2D | other | 11.00 | download | — | — | — | — | MMLU 76.84%; MMLU_base 80.0% | other or no reference | L118 @ main |
| dealignai/MXFP8-CRACK [L118]publisher comparison build | 6D | MXFP8 | 24.00 | download | — | — | — | — | MMLU 89.12%; HarmBench-240 100.0% | other or no reference | L118 @ main |
| dealignai/MXFP8-CRACK [L117]publisher comparison build | 4D | MXFP8 | 17.00 | download | — | — | — | — | MMLU 87.72%; HarmBench-240 100.0% | other or no reference | L117 @ main |
| dealignai/MXFP8-CRACK other repo variant [L116]publisher comparison build | 2D | MXFP8 | 11.00 | download | — | — | — | — | MMLU 76.84%; HarmBench-240 100.0% | other or no reference | L116 @ main |
| Derpyhue/abliterated-2.55bpw-SC-SH median c19, p90 c20; perplexity column left out [L48]publisher comparison build | 2.65 H4 -sh | other | 8.8 | text weights (excl. vision, MTP) | 0.0377 | — | — | — | — | vs BF16 | L48 @ main |
| Derpyhue/abliterated-2.55bpw-SC-SH median c14, p90 c15; perplexity column left out [L47]publisher comparison build | 2.60 H4 -sh | other | 8.6 | text weights (excl. vision, MTP) | 0.0417 | — | — | — | — | vs BF16 | L47 @ main |
| Derpyhue/abliterated-2.60bpw-SC-SH unquantized reference row [L45]publisher comparison build | Abliterated BF16 (reference) | other | 51.22 | text weights (excl. vision, MTP) | — | — | 1.3418 | — | — | vs BF16 | L45 @ main |
| Derpyhue/abliterated-2.60bpw-SC-SH PPL cell reads 1.3986 (+4.2%) [L48]bitrate 2.65 is from row label onlypublisher comparison build | 2.65 H4 -sh | other | 8.8 | text weights (excl. vision, MTP) | 0.0377 | — | 1.3986 (+4.2%) | — | — | vs BF16 | L48 @ main |
| Derpyhue/abliterated-2.60bpw-SC-SH PPL cell reads 1.4098 (+5.1%) [L46]bitrate 2.55 is from row label onlypublisher comparison build | 2.55 H4 -sh | other | 8.4 | text weights (excl. vision, MTP) | 0.0442 | — | 1.4098 (+5.1%) | — | — | vs BF16 | L46 @ main |
| Derpyhue/abliterated-2.65bpw-SC-SH Perplexity (increase) ambiguous, not mapped [L47]publisher comparison build | 2.60 H4 -sh | other | 8.6 | text weights (excl. vision, MTP) | 0.0417 | — | — | — | — | vs BF16 | L47 @ main |
| Derpyhue/abliterated-2.65bpw-SC-SH Perplexity (increase) ambiguous, not mapped [L46]publisher comparison build | 2.55 H4 -sh | other | 8.4 | text weights (excl. vision, MTP) | 0.0442 | — | — | — | — | vs BF16 | L46 @ main |
| Derpyhue/exl3-2.65bpw-SC publisher comparison build | 2.80 H4 | EXL3 | 12.57 | download | 0.025571 | — | 1.3690 | — | — | vs BF16 | L48 @ main |
| Derpyhue/exl3-2.65bpw-SC publisher comparison build | 2.80 H4 -sh | EXL3 | 12.57 | download | 0.025098 | — | 1.3688 | — | — | vs BF16 | L49 @ main |
| Derpyhue/exl3-2.65bpw-SC publisher comparison build | 2.75 H4 | EXL3 | 12.41 | download | 0.027894 | — | 1.3725 | — | — | vs BF16 | L46 @ main |
| Derpyhue/exl3-2.65bpw-SC publisher comparison build | 2.75 H4 -sh | EXL3 | 12.41 | download | 0.026945 | — | 1.3685 | — | — | vs BF16 | L47 @ main |
| Derpyhue/exl3-2.65bpw-SC publisher comparison build | 2.70 H4 | EXL3 | 12.26 | download | 0.030083 | — | 1.3746 | — | — | vs BF16 | L44 @ main |
| Derpyhue/exl3-2.65bpw-SC publisher comparison build | 2.70 H4 -sh | EXL3 | 12.26 | download | 0.030040 | — | 1.3748 | — | — | vs BF16 | L45 @ main |
| Derpyhue/exl3-2.65bpw-SC publisher comparison build | 2.67 H4 -sh | EXL3 | 12.17 | download | 0.031061 | — | 1.3748 | — | — | vs BF16 | L43 @ main |
| Derpyhue/exl3-2.65bpw-SC separate repo variant [L16]publisher comparison build | 2.65 H4 -sh | EXL3 | 12.11 | download | 0.032541 | — | 1.3739 | — | — | vs BF16 | L42 @ main |
| Derpyhue/exl3-2.65bpw-SC-SH publisher comparison build | 2.80 H4 | EXL3 | 12.57 | download | 0.025571 | — | 1.3690 | — | — | vs BF16 | L49 @ main |
| Derpyhue/exl3-2.65bpw-SC-SH publisher comparison build | 2.80 H4 -sh | EXL3 | 12.57 | download | 0.025098 | — | 1.3688 | — | — | vs BF16 | L50 @ main |
| Derpyhue/exl3-2.65bpw-SC-SH publisher comparison build | 2.75 H4 | EXL3 | 12.41 | download | 0.027894 | — | 1.3725 | — | — | vs BF16 | L47 @ main |
| Derpyhue/exl3-2.65bpw-SC-SH publisher comparison build | 2.75 H4 -sh | EXL3 | 12.41 | download | 0.026945 | — | 1.3685 | — | — | vs BF16 | L48 @ main |
| Derpyhue/exl3-2.65bpw-SC-SH publisher comparison build | 2.70 H4 | EXL3 | 12.26 | download | 0.030083 | — | 1.3746 | — | — | vs BF16 | L45 @ main |
| Derpyhue/exl3-2.65bpw-SC-SH publisher comparison build | 2.70 H4 -sh | EXL3 | 12.26 | download | 0.030040 | — | 1.3748 | — | — | vs BF16 | L46 @ main |
| Derpyhue/exl3-2.65bpw-SC-SH publisher comparison build | 2.67 H4 -sh | EXL3 | 12.17 | download | 0.031061 | — | 1.3748 | — | — | vs BF16 | L44 @ main |
| Derpyhue/exl3-2.65bpw-SC-SH publisher comparison build | 2.65 H4 | EXL3 | 12.11 | download | 0.032903 | — | 1.3795 | — | — | vs BF16 | L42 @ main |
| Derpyhue/exl3-2.67bpw-SC-SH [L49]publisher comparison build | 2.80 H4 | EXL3 | 12.57 | download | 0.025571 | — | 1.3690 | — | — | vs BF16 | L49 @ main |
| Derpyhue/exl3-2.67bpw-SC-SH [L50]publisher comparison build | 2.80 H4 -sh | EXL3 | 12.57 | download | 0.025098 | — | 1.3688 | — | — | vs BF16 | L50 @ main |
| Derpyhue/exl3-2.67bpw-SC-SH [L47]publisher comparison build | 2.75 H4 | EXL3 | 12.41 | download | 0.027894 | — | 1.3725 | — | — | vs BF16 | L47 @ main |
| Derpyhue/exl3-2.67bpw-SC-SH [L48]publisher comparison build | 2.75 H4 -sh | EXL3 | 12.41 | download | 0.026945 | — | 1.3685 | — | — | vs BF16 | L48 @ main |
| Derpyhue/exl3-2.67bpw-SC-SH [L45]publisher comparison build | 2.70 H4 | EXL3 | 12.26 | download | 0.030083 | — | 1.3746 | — | — | vs BF16 | L45 @ main |
| Derpyhue/exl3-2.67bpw-SC-SH [L46]publisher comparison build | 2.70 H4 -sh | EXL3 | 12.26 | download | 0.030040 | — | 1.3748 | — | — | vs BF16 | L46 @ main |
| Derpyhue/exl3-2.67bpw-SC-SH [L42]publisher comparison build | 2.65 H4 | EXL3 | 12.11 | download | 0.032903 | — | 1.3795 | — | — | vs BF16 | L42 @ main |
| Derpyhue/exl3-2.67bpw-SC-SH [L43]publisher comparison build | 2.65 H4 -sh | EXL3 | 12.11 | download | 0.032541 | — | 1.3739 | — | — | vs BF16 | L43 @ main |
| Derpyhue/exl3-2.70bpw-SC [L48]publisher comparison build | 2.80 H4 | EXL3 | 12.57 | download | 0.025571 | — | 1.3690 | — | — | vs BF16 | L48 @ main |
| Derpyhue/exl3-2.70bpw-SC [L49]publisher comparison build | 2.80 H4 -sh | EXL3 | 12.57 | download | 0.025098 | — | 1.3688 | — | — | vs BF16 | L49 @ main |
| Derpyhue/exl3-2.70bpw-SC [L46]publisher comparison build | 2.75 H4 | EXL3 | 12.41 | download | 0.027894 | — | 1.3725 | — | — | vs BF16 | L46 @ main |
| Derpyhue/exl3-2.70bpw-SC [L47]publisher comparison build | 2.75 H4 -sh | EXL3 | 12.41 | download | 0.026945 | — | 1.3685 | — | — | vs BF16 | L47 @ main |
| Derpyhue/exl3-2.70bpw-SC [L45]publisher comparison build | 2.70 H4 -sh | EXL3 | 12.26 | download | 0.030040 | — | 1.3748 | — | — | vs BF16 | L45 @ main |
| Derpyhue/exl3-2.70bpw-SC [L43]publisher comparison build | 2.67 H4 -sh | EXL3 | 12.17 | download | 0.031061 | — | 1.3748 | — | — | vs BF16 | L43 @ main |
| Derpyhue/exl3-2.70bpw-SC [L41]publisher comparison build | 2.65 H4 | EXL3 | 12.11 | download | 0.032903 | — | 1.3795 | — | — | vs BF16 | L41 @ main |
| Derpyhue/exl3-2.70bpw-SC [L42]publisher comparison build | 2.65 H4 -sh | EXL3 | 12.11 | download | 0.032541 | — | 1.3739 | — | — | vs BF16 | L42 @ main |
| Derpyhue/exl3-2.70bpw-SC-SH [L49]publisher comparison build | 2.80 H4 | EXL3 | 12.57 | download | 0.025571 | — | 1.3690 | — | — | vs BF16 | L49 @ main |
| Derpyhue/exl3-2.70bpw-SC-SH [L50]publisher comparison build | 2.80 H4 -sh | EXL3 | 12.57 | download | 0.025098 | — | 1.3688 | — | — | vs BF16 | L50 @ main |
| Derpyhue/exl3-2.70bpw-SC-SH [L47]publisher comparison build | 2.75 H4 | EXL3 | 12.41 | download | 0.027894 | — | 1.3725 | — | — | vs BF16 | L47 @ main |
| Derpyhue/exl3-2.70bpw-SC-SH [L48]publisher comparison build | 2.75 H4 -sh | EXL3 | 12.41 | download | 0.026945 | — | 1.3685 | — | — | vs BF16 | L48 @ main |
| Derpyhue/exl3-2.70bpw-SC-SH [L45]publisher comparison build | 2.70 H4 | EXL3 | 12.26 | download | 0.030083 | — | 1.3746 | — | — | vs BF16 | L45 @ main |
| Derpyhue/exl3-2.70bpw-SC-SH [L44]publisher comparison build | 2.67 H4 -sh | EXL3 | 12.17 | download | 0.031061 | — | 1.3748 | — | — | vs BF16 | L44 @ main |
| Derpyhue/exl3-2.70bpw-SC-SH [L42]publisher comparison build | 2.65 H4 | EXL3 | 12.11 | download | 0.032903 | — | 1.3795 | — | — | vs BF16 | L42 @ main |
| Derpyhue/exl3-2.70bpw-SC-SH [L43]publisher comparison build | 2.65 H4 -sh | EXL3 | 12.11 | download | 0.032541 | — | 1.3739 | — | — | vs BF16 | L43 @ main |
| Derpyhue/exl3-2.75bpw-SC bpw 2.80 only in row label [L48]publisher comparison build | 2.80 H4 | EXL3 | 12.57 | download | 0.025571 | — | 1.3690 | — | — | vs BF16 | L48 @ main |
| Derpyhue/exl3-2.75bpw-SC bpw 2.80 only in row label [L49]publisher comparison build | 2.80 H4 -sh | EXL3 | 12.57 | download | 0.025098 | — | 1.3688 | — | — | vs BF16 | L49 @ main |
| Derpyhue/exl3-2.75bpw-SC separate repo, scaled-Hessian variant [L16,L47]publisher comparison build | 2.75 H4 -sh | EXL3 | 12.41 | download | 0.026945 | — | 1.3685 | — | — | vs BF16 | L47 @ main |
| Derpyhue/exl3-2.75bpw-SC bpw 2.70 only in row label [L44]publisher comparison build | 2.70 H4 | EXL3 | 12.26 | download | 0.030083 | — | 1.3746 | — | — | vs BF16 | L44 @ main |
| Derpyhue/exl3-2.75bpw-SC bpw 2.70 only in row label [L45]publisher comparison build | 2.70 H4 -sh | EXL3 | 12.26 | download | 0.030040 | — | 1.3748 | — | — | vs BF16 | L45 @ main |
| Derpyhue/exl3-2.75bpw-SC bpw 2.67 only in row label [L43]publisher comparison build | 2.67 H4 -sh | EXL3 | 12.17 | download | 0.031061 | — | 1.3748 | — | — | vs BF16 | L43 @ main |
| Derpyhue/exl3-2.75bpw-SC bpw 2.65 only in row label [L41]publisher comparison build | 2.65 H4 | EXL3 | 12.11 | download | 0.032903 | — | 1.3795 | — | — | vs BF16 | L41 @ main |
| Derpyhue/exl3-2.75bpw-SC bpw 2.65 only in row label [L42]publisher comparison build | 2.65 H4 -sh | EXL3 | 12.11 | download | 0.032541 | — | 1.3739 | — | — | vs BF16 | L42 @ main |
| Derpyhue/exl3-2.75bpw-SC-SH publisher comparison build | 2.80 H4 | EXL3 | 12.57 | download | 0.025571 | — | 1.3690 | — | — | vs BF16 | L49 @ main |
| Derpyhue/exl3-2.75bpw-SC-SH publisher comparison build | 2.80 H4 -sh | EXL3 | 12.57 | download | 0.025098 | — | 1.3688 | — | — | vs BF16 | L50 @ main |
| Derpyhue/exl3-2.75bpw-SC-SH publisher comparison build | 2.75 H4 | EXL3 | 12.41 | download | 0.027894 | — | 1.3725 | — | — | vs BF16 | L47 @ main |
| Derpyhue/exl3-2.75bpw-SC-SH publisher comparison build | 2.70 H4 | EXL3 | 12.26 | download | 0.030083 | — | 1.3746 | — | — | vs BF16 | L45 @ main |
| Derpyhue/exl3-2.75bpw-SC-SH publisher comparison build | 2.70 H4 -sh | EXL3 | 12.26 | download | 0.030040 | — | 1.3748 | — | — | vs BF16 | L46 @ main |
| Derpyhue/exl3-2.75bpw-SC-SH publisher comparison build | 2.67 H4 -sh | EXL3 | 12.17 | download | 0.031061 | — | 1.3748 | — | — | vs BF16 | L44 @ main |
| Derpyhue/exl3-2.75bpw-SC-SH Median/p90 KLD columns not mapped (no key) [L42]; perplexity ~1.37 is unusually low, corpus not statedpublisher comparison build | 2.65 H4 | EXL3 | 12.11 | download | 0.032903 | — | 1.3795 | — | — | vs BF16 | L42 @ main |
| Derpyhue/exl3-2.75bpw-SC-SH Size is full repo footprint [L38]publisher comparison build | 2.65 H4 -sh | EXL3 | 12.11 | download | 0.032541 | — | 1.3739 | — | — | vs BF16 | L43 @ main |
| Derpyhue/exl3-2.80bpw-SC separate SH repopublisher comparison build | 2.80 H4 -sh | EXL3 | 12.57 | download | 0.025098 | — | 1.3688 | — | kld_median 0.001232; kld_p90 0.073292 | vs BF16 | L49 @ main |
| Derpyhue/exl3-2.80bpw-SC publisher comparison build | 2.75 H4 | EXL3 | 12.41 | download | 0.027894 | — | 1.3725 | — | kld_median 0.001434; kld_p90 0.081578 | vs BF16 | L46 @ main |
| Derpyhue/exl3-2.80bpw-SC publisher comparison build | 2.75 H4 -sh | EXL3 | 12.41 | download | 0.026945 | — | 1.3685 | — | kld_median 0.001279; kld_p90 0.077931 | vs BF16 | L47 @ main |
| Derpyhue/exl3-2.80bpw-SC publisher comparison build | 2.70 H4 | EXL3 | 12.26 | download | 0.030083 | — | 1.3746 | — | kld_median 0.001484; kld_p90 0.087552 | vs BF16 | L44 @ main |
| Derpyhue/exl3-2.80bpw-SC publisher comparison build | 2.70 H4 -sh | EXL3 | 12.26 | download | 0.030040 | — | 1.3748 | — | kld_median 0.001454; kld_p90 0.086279 | vs BF16 | L45 @ main |
| Derpyhue/exl3-2.80bpw-SC publisher comparison build | 2.67 H4 -sh | EXL3 | 12.17 | download | 0.031061 | — | 1.3748 | — | kld_median 0.001562; kld_p90 0.089724 | vs BF16 | L43 @ main |
| Derpyhue/exl3-2.80bpw-SC publisher comparison build | 2.65 H4 | EXL3 | 12.11 | download | 0.032903 | — | 1.3795 | — | kld_median 0.001696; kld_p90 0.095268 | vs BF16 | L41 @ main |
| Derpyhue/exl3-2.80bpw-SC publisher comparison build | 2.65 H4 -sh | EXL3 | 12.11 | download | 0.032541 | — | 1.3739 | — | kld_median 0.001560; kld_p90 0.092880 | vs BF16 | L42 @ main |
| Derpyhue/exl3-2.80bpw-SC-SH non -sh sibling; SC build without scaled Hessian [L16,L49]publisher comparison build | 2.80 H4 | EXL3 | 12.57 | download | 0.025571 | — | 1.3690 | — | — | vs BF16 | L24 @ main |
| Derpyhue/exl3-2.80bpw-SC-SH [L47]publisher comparison build | 2.75 H4 | EXL3 | 12.41 | download | 0.027894 | — | 1.3725 | — | — | vs BF16 | L47 @ main |
| Derpyhue/exl3-2.80bpw-SC-SH [L48]publisher comparison build | 2.75 H4 -sh | EXL3 | 12.41 | download | 0.026945 | — | 1.3685 | — | — | vs BF16 | L48 @ main |
| Derpyhue/exl3-2.80bpw-SC-SH [L45]publisher comparison build | 2.70 H4 | EXL3 | 12.26 | download | 0.030083 | — | 1.3746 | — | — | vs BF16 | L45 @ main |
| Derpyhue/exl3-2.80bpw-SC-SH [L46]publisher comparison build | 2.70 H4 -sh | EXL3 | 12.26 | download | 0.030040 | — | 1.3748 | — | — | vs BF16 | L46 @ main |
| Derpyhue/exl3-2.80bpw-SC-SH [L44]publisher comparison build | 2.67 H4 -sh | EXL3 | 12.17 | download | 0.031061 | — | 1.3748 | — | — | vs BF16 | L44 @ main |
| Derpyhue/exl3-2.80bpw-SC-SH 2.65 bpw from row label [L42]publisher comparison build | 2.65 H4 | EXL3 | 12.11 | download | 0.032903 | — | 1.3795 | — | — | vs BF16 | L42 @ main |
| Derpyhue/exl3-2.80bpw-SC-SH [L43]publisher comparison build | 2.65 H4 -sh | EXL3 | 12.11 | download | 0.032541 | — | 1.3739 | — | — | vs BF16 | L43 @ main |
| eaddario/GGUF baseline [L110]publisher comparison build | Qwen3.8-27B-F16 | GGUF | 51.00 | download | — | — | 6.669665 ±0.042605 | — | — | other or no reference | L110 @ main |
| EliovpAI/W3Rot-INT3-Paiton-RDNA4 GSM8K/HumanEval/MMLU-Pro 0-shot from 26 Sep table [L159-161], greedy thinking offpublisher comparison build | Previous weights (7a2e3d70, W3A4) | other | — | — | — | — | — | kld:decode 0.0851; kld:prefill 0.1074; top1:decode 88.66; top1:prefill 87.36; kld:long_decode 0.0806; kld:long_prefill 0.0975; kld:r9700_served 0.1732; top1:r9700_served 86.90 %; GSM8K 95.30 %; HumanEval 93.90 % | vs BF16 | L124 @ main | |
| EliovpAI/W3Rot-INT3-Paiton-RDNA4 [L159-162]; Needle 100% for MXFP4; model memory 19.18 GiB vs 15.89 GiB in prose [L195-196]publisher comparison build | MXFP4 image (previous publisher build) | other | — | — | — | — | — | GSM8K 95.68 %; HumanEval 95.12 %; Needle_61440 100 % | vs BF16 | L159 @ main | |
| empero-ai/Ridge measured by tooltdrepo name from update note [L106-107] | Ridge-3.7bpw (empero-ai) | 11.73 | download | 0.118430 | 85.907% | 7.547496 | — | HumanEval 0.933; HumanEval+ 0.896; MBPP 0.902; MBPP+ 0.765 | vs BF16 | L42 @ main | |
| FedericoSciuca/depthaware-GGUF reference row [L35]publisher comparison build | reference (Q8_0) | GGUF | 29.05 | download | — | — | 5.5950 | — | — | vs a Q8_0 | L35 @ main |
| FedericoSciuca/depthaware-GGUF byte-identical size to this model [L37]excess +0.8186 c9 absolutepublisher comparison build | control 13 blocks spread evenly | GGUF | 12.48 | download | — | — | 6.4136 | — | — | vs a Q8_0 | L37 @ main |
| Freaksterz/SmoothQuant-W8A8-INT8 floor-free KLD c15, PPL delta-ln c17 [L31]; breaks DFlash2 [L16]publisher comparison build | v2-rotated W8A8 | INT8 | — | 0.01098 | 95.9% | — | — | — | vs BF16 | L31 @ main | |
| Freaksterz/SmoothQuant-W8A8-INT8 same table [L32]publisher comparison build | v1-smoothquant-rtn W8A8 | INT8 | — | 0.01414 | 95.3% | — | — | — | vs BF16 | L32 @ main | |
| Frosty40/GGUF size shown is the -MTP file [L30]tail is 99.9th percentile [L47]publisher comparison build | Q6_K-OQ8(-MTP) | GGUF | 21.30 | download | 0.00170 kld:prose | — | — | — | top1:prose 97.79; kld:code 0.00117; top1:code 99.46; kld:agentic 0.00137; top1:agentic 99.36 | vs BF16 | L30 @ main |
| GestaltLabs/EXL3-11.5GB publisher comparison build | Prior Release (Pilot Calibration) | EXL3 | — | — | — | — | — | kld:development 0.18597; kld:acceptance 0.26780; kld:combined 0.22689 | other or no reference | L129 @ main | |
| GestaltLabs/EXL3-11.5GB publisher comparison build | Pilot 8K Candidate (64 rows) | EXL3 | — | — | — | — | — | kld:development 0.23696; kld:acceptance 0.33439; kld:combined 0.28568 | other or no reference | L130 @ main | |
| ggml-org (repo name truncated in card digest) measured by fraQtllarger build, no NextN block [L41] | ggml-org second | 18.97 | download | — | — | — | — | kld:code/math 0.19074; kld:general 0.24274; GSM8K 94.0% (188/200); MATH-500 87.4% (437/500) | vs a Q8_0 | L36 @ main | |
| ggml-org (repo not named on card) measured by AtomicChat[L216] | Q4_K_M (ggml-org) | 19.00 | download | 0.01470 | — | — | — | — | vs BF16 | L216 @ main | |
| gittensor-model-hub/NVFP4-RTX5090 measured by malaiwahresident 18.77 GiB measured [L486]; shard-0 [L464] | gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090 | NVFP4 | — | 0.062163 | 89.85 % | — | — | — | vs BF16 | L464 @ main | |
| gittensor-model-hub/NVFP4-RTX5090 measured by malaiwahshard-0 of v5 suite [L473]measured resident 18.77 GiB appears only in prose [L487] | gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090 | NVFP4 | — | 0.062163 | 89.85 % | — | — | — | vs BF16 | L473 @ main | |
| gittensor-model-hub/NVFP4-RTX5090 measured by malaiwah[L504]; resident 18.77 GiB in prose [L522] | gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090 shard0 | NVFP4 | — | 0.062163 | 89.85 % | — | — | — | other or no reference | L504 @ main | |
| glyd/kestrel reference model; GSM8K bf16 column served by vLLM [L52]publisher comparison build | bf16 (original) | other | 53.80 | download | 0 | — | — | — | GSM8K 96.7% | vs BF16 | L26 @ main |
| glyd/kestrel publisher comparison build | penguin | other | 37.00 | download | 0 | — | — | — | — | vs BF16 | L28 @ main |
| glyd/kestrel publisher comparison build | swift | other | 18.60 | download | 0.0139 | — | — | — | — | vs BF16 | L30 @ main |
| glyd/penguin publisher comparison build | bf16 (original) | other | 53.80 | download | 0 | — | — | — | — | vs BF16 | L31 @ main |
| glyd/penguin Glyd's other build [L34]publisher comparison build | kestrel | other | 22.00 | download | 0.00477 | — | — | — | — | vs BF16 | L34 @ main |
| glyd/penguin Glyd's other build [L35]publisher comparison build | swift | other | 18.60 | download | 0.0139 | — | — | — | — | vs BF16 | L35 @ main |
| glyd/swift reference row [L26]; GSM8K from vLLM-served bf16 [L52]publisher comparison build | bf16 (original) | other | 53.80 | download | 0 | — | — | — | — | vs BF16 | L26 @ main |
| glyd/swift KL listed as 0 [L28]publisher comparison build | penguin | other | 37.00 | download | 0 | — | — | — | — | vs BF16 | L28 @ main |
| glyd/swift [L29]publisher comparison build | kestrel | other | 22.00 | download | 0.00477 | — | — | — | — | vs BF16 | L29 @ main |
| gratex/W4A16-g128-sym-GPTQ unquantized reference [L109]publisher comparison build | BF16 base | GPTQ | 54.00 | download | — | — | 9.6402 | — | — | vs BF16 | L109 @ main |
| guideboardlabs/Q3-DOWN-XS-GGUF overflows 8 GB [L50]publisher comparison build | Q3-DOWN-UP-XS (ceiling) | GGUF | — | — | — | — | — | ppl:wikitext-2 8.1278 | other or no reference | L45 @ main | |
| guideboardlabs/Q3-DOWN-XS-GGUF [L47]publisher comparison build | Q3-UP-XS | GGUF | — | — | — | — | — | ppl:wikitext-2 8.3272 | other or no reference | L47 @ main | |
| guideboardlabs/Q3-DOWN-XS-GGUF [L48]publisher comparison build | Q3-GATE-PROTECT (old prod) | GGUF | — | — | — | — | — | ppl:wikitext-2 8.3561 | other or no reference | L48 @ main | |
| HangGlidersRule/Darkstar-Base-ModelOpt-W4A16-NVFP4-Mixed-FP8 157/198 BF16 control vs 153/198 quant, a 4-question delta [L44]publisher comparison build | upstream BF16 control | NVFP4 | — | — | — | — | — | GPQA-Diamond_correct_of_198 157 | vs BF16 | L44 @ main | |
| HermiHg/mmproj-Q5_K-MIX-GGUF reference projector [L44-57]publisher comparison build | BF16 reference | GGUF | — | — | — | — | — | TextVQA 0.8204; OCRBench-v2 0.6623; GQA 0.6280; ScienceQA 0.9450; MMStar 0.7140; ChartQA 0.8520; BLINK jigsaw 0.5200; BLINK relative depth 0.9113; BLINK visual similarity 0.9037; BLINK counting 0.8167; BLINK object localization 0.8033; Overall 0.7493 | vs BF16 | L46 @ main | |
| Inferact/NVFP4 measured by QUASAR-QAT304/496 linears NVFP4 [L85] | Inferact NVFP4 | NVFP4 | 26.40 | download | — | — | — | — | GPQA-Diamond 87.63; AIME'26 96.67 | vs BF16 | L85 @ main |
| inferencerlabs/MLX-Q9 bpw from row label [L34]publisher comparison build | Q3.5 | MLX | — | — | 43.45% | 168.0 | — | — | vs BF16 | L34 @ main | |
| inferencerlabs/MLX-Q9 bpw from row label [L35]publisher comparison build | Q4.5 | MLX | — | — | 91.65% | 1.33593 | — | — | vs BF16 | L35 @ main | |
| inferencerlabs/MLX-Q9 bpw from row label [L36]publisher comparison build | Q5.5 | MLX | — | — | 95.05% | 1.23437 | — | — | vs BF16 | L36 @ main | |
| inferencerlabs/MLX-Q9 bpw from row label [L37]publisher comparison build | Q6.5 | MLX | — | — | 96.65% | 1.21875 | — | — | vs BF16 | L37 @ main | |
| inferencerlabs/MLX-Q9 bpw from row label [L38]publisher comparison build | Q8.5 | MLX | — | — | 97.65% | 1.21875 | — | — | vs BF16 | L38 @ main | |
| ionos/GPTQ-Int4 measured by card author, paired [L35-36]publisher comparison build | Base (BF16) | GPTQ | — | — | — | — | — | OmniDocBench_reading_order 92.6%; OmniDocBench_display_formula 72.7%; MathVista 85.7%; MMMU 81.0%; V*Bench 88.7%; MMLU-Pro 86.3%; MMLU-DE 82.7%; BFCL 57.7%; IFEval 40.1% | vs BF16 | L40 @ main | |
| ISTA-DASLab/GSQ-RCO-GGUF measured by jdominick05repo id taken from related-quants link [L145]; table row spells the org IST-DASLab [L47] | IST-DASLab GSQ-RCO IQ3_S | GGUF | 11.77 | download | 0.0645 ± 0.0028 | 92.99% | — | +5.2% | — | vs a Q8_0 | L47 @ main |
| IsValorum/VAL-APEX-I-MiniPlus-V3-GGUF Size 54.65 GB (50.90 GiB) [L72]; PPL has large +/- 0.4531publisher comparison build | BF16 Base baseline (unmodified) | GGUF | 54.65 | download | — | — | 5.7942 ± 0.4531 | — | — | vs BF16 | L72 @ main |
| IsValorum/VAL-APEX-I-NanoPlus-V3-GGUF reference row [L72]publisher comparison build | BF16 Base baseline (unmodified) | GGUF | 54.65 | download | — | — | — | — | ppl:wikitext2 5.7942 ± 0.4531 | vs BF16 | L72 @ main |
| jaakkomoller/q3mix-NInfer pure-Q4 16 GB profile [L88]publisher comparison build | q4 | other | — | — | — | — | — | LiveCodeBench_v6_175 41/175; quality_gate_35 28/35 | other or no reference | L180 @ main | |
| JANGQ-AI/JANGH2 Sibling bundle from same publisher, median KL [L66-79]publisher comparison build | JANG_4D by topic (held-out text) | other | — | — | — | — | — | kld:coding 0.0002; top1:coding 98.4 %; kld:math 0.0001; top1:math 98.7 %; kld:office_finance 0.0001; top1:office_finance 97.7 %; kld:topic_reasoning 0.0001; top1:topic_reasoning 94.5 %; kld:science 0.0010; top1:science 98.0 %; kld:long_context 0.0007; top1:long_context 96.6 %; kld:cybersecurity 0.0008; top1:cybersecurity 94.9 %; kld:networking 0.0011; top1:networking 93.7 %; kld:terminal_shell 0.0013; top1:terminal_shell 92.5 %; kld:agentic_tool_calls 0.0026; top1:agentic_tool_calls 89.8 %; kld:creative_writing 0.0028; top1:creative_writing 96.5 %; kld:multi_turn_agentic 0.0014; top1:multi_turn_agentic 92.1 % | vs BF16 | L68 @ main | |
| JANGQ-AI/JANGH2 Sibling bundle from same publisher, median KL [L83-91]publisher comparison build | JANG_4D by topic (own bf16 reasoning traces) | other | — | — | — | — | — | kld:traces_coding 0.00007; top1:traces_coding 98.4 %; kld:traces_math 0.00001; top1:traces_math 98.6 %; kld:traces_office_finance 0.00000; top1:traces_office_finance 98.5 %; kld:traces_science 0.00010; top1:traces_science 98.1 %; kld:traces_networking 0.00000; top1:traces_networking 98.7 %; kld:traces_cybersecurity 0.00135; top1:traces_cybersecurity 97.3 %; kld:traces_agentic_tool_calls 0.00093; top1:traces_agentic_tool_calls 89.7 % | vs BF16 | L85 @ main | |
| JANGQ-AI/JANG_4D Sibling bundle by same publisher; median KL [L66]; size not stated on cardpublisher comparison build | JANGH2 (by topic, held-out text set) | other | — | — | — | — | — | kld:coding 0.0092; top1:coding 89.5 %; kld:math 0.0068; top1:math 90.8 %; kld:office_finance 0.0087; top1:office_finance 86.9 %; kld:reasoning 0.0081; top1:reasoning 81.0 %; kld:science 0.0460; top1:science 87.1 %; kld:long_context_attention_recall 0.0446; top1:long_context_attention_recall 80.4 %; kld:cybersecurity 0.0368; top1:cybersecurity 80.5 %; kld:networking 0.0244; top1:networking 81.9 %; kld:terminal_shell 0.0738; top1:terminal_shell 73.5 %; kld:agentic_tool_calls 0.0975; top1:agentic_tool_calls 71.2 %; kld:creative_writing 0.1989; top1:creative_writing 74.2 %; kld:multi_turn_agentic 0.0615; top1:multi_turn_agentic 75.3 % | vs BF16 | L70 @ main | |
| JANGQ-AI/JANG_4D Sibling bundle by same publisher; median KL [L83-85]publisher comparison build | JANGH2 (own bf16 reasoning traces, by topic) | other | — | — | — | — | — | kld:coding 0.00508; top1:coding 89.4 %; kld:math 0.00061; top1:math 91.6 %; kld:office_finance 0.00007; top1:office_finance 91.2 %; kld:science 0.00779; top1:science 87.8 %; kld:networking 0.00009; top1:networking 92.8 %; kld:cybersecurity 0.08516; top1:cybersecurity 81.1 %; kld:agentic_tool_calls 0.03968; top1:agentic_tool_calls 72.6 % | vs BF16 | L87 @ main | |
| jcbtc/CIRU-ActiveFPX-PromptForge previous release, matched measure [L53]publisher comparison build | Public v2 | other | — | 0.036162 | 90.637% | — | 1.027876 | — | vs BF16 | L55 @ main | |
| jclyons52/Unsloth-Dynamic-v3-mlx author's own re-encoding for comparison [L51]publisher comparison build | same checkpoint, Unsloth Dynamic 2.0 static per-tensor table (5.27 bpw) | MLX | 17.00 | download | — | — | — | — | ppl:wikitext-2 13.445 | reference not stated | L51 @ main |
| jdominick05/24GB-262k-GGUF retired build [L45]publisher comparison build | R1-mix (retired) | GGUF | 16.95 | download | 0.0191 ± 0.0015 | 94.96% | — | +0.8% | — | vs a Q8_0 | L45 @ main |
| jdominick05/24GB-262k-GGUF retired build [L44]publisher comparison build | R2-mix (retired) | GGUF | 15.39 | download | 0.0199 ± 0.0012 | 95.08% | — | +0.5% | — | vs a Q8_0 | L44 @ main |
| Jon-Nielsen/exl3-4.00bpw Sibling build by same publisher; VRAM peak [L21]publisher comparison build | EXL3 6.00bpw | EXL3 | 30.00 | GPU weights | — | — | — | — | ppl:40K_tech 5.07 | reference not stated | L21 @ main |
| Jon-Nielsen/exl3-6.00bpw Sibling repo Jon-Nielsen/Qwen3.8-27B-exl3-4.00bpw [L11,L20]publisher comparison build | 4.00bpw | EXL3 | 24.00 | GPU weights | — | — | — | — | ppl:40K_tech 5.25 | reference not stated | L20 @ main |
| JRamirez-UAB/EXL3-3.85bpw-24GB publisher's sibling GPTQ repo [L88]publisher comparison build | GPTQ W4A16 g64 + INT8 embeddings | EXL3 | 16.91 | download | 0.0175 | — | 1.6203 | — | kld:public_trace 0.0171; ppl:public_trace 1.3568 | vs BF16 | L88 @ main |
| JRamirez-UAB/EXL3-3.85bpw-24GB publisher's sibling GPTQ repo [L87]publisher comparison build | GPTQ W4A16 + INT4 embeddings | EXL3 | 15.26 | download | 0.0400 | — | 1.6631 | — | kld:public_trace 0.0312; ppl:public_trace 1.3779 | vs BF16 | L87 @ main |
| JRamirez-UAB/GPTQ-W4A16-embed-int4-24GB same publisher, same conditions [L87]publisher comparison build | GPTQ W4A16 g64 + INT8 embeddings | GPTQ | 16.91 | download | — | — | — | — | kld:internal 0.0175; kld:public 0.0171; ppl:internal 1.6203; ppl:public 1.3568 | vs BF16 | L92 @ main |
| JRamirez-UAB/GPTQ-W4A16-embed-int4-24GB publisher's sibling repo JRamirez-UAB/Qwen3.8-27B-EXL3-3.85bpw-24GB [L93]publisher comparison build | EXL3 3.85 bpw (main) | GPTQ | 14.62 | download | — | — | — | — | kld:internal 0.0073; kld:public 0.0067; ppl:internal 1.5976; ppl:public 1.3430 | vs BF16 | L93 @ main |
| JRamirez-UAB/GPTQ-W4A16-embed-int4-24GB branch of EXL3-3.85bpw-24GB repo [L94]publisher comparison build | EXL3 3.70 bpw (3.70bpw branch) | GPTQ | 14.16 | download | — | — | — | — | kld:internal 0.0089; kld:public 0.0085; ppl:internal 1.5999; ppl:public 1.3466 | vs BF16 | L94 @ main |
| JRamirez-UAB/GPTQ-W4A16-embed-int4-24GB BF16 perplexity given in column header [L89]publisher comparison build | BF16 reference | GPTQ | — | — | — | — | — | ppl:internal BF16 1.5860 | vs BF16 | L89 @ main | |
| JRamirez-UAB/GPTQ-W4A16-g64-embed-int8-24GB [L82]publisher comparison build | GPTQ W4A16 + INT4 embeddings | GPTQ | 15.26 | download | — | — | — | — | kld:internal 0.0400; kld:public 0.0312; ppl:internal 1.6631; ppl:public 1.3779 | vs BF16 | L82 @ main |
| JRamirez-UAB/GPTQ-W4A16-g64-embed-int8-24GB [L84]publisher comparison build | EXL3 3.85 bpw | GPTQ | 14.62 | download | — | — | — | — | kld:internal 0.0073; kld:public 0.0067; ppl:internal 1.5976; ppl:public 1.3430 | vs BF16 | L84 @ main |
| JRamirez-UAB/GPTQ-W4A16-g64-embed-int8-24GB [L85]publisher comparison build | EXL3 3.70 bpw | GPTQ | 14.16 | download | — | — | — | — | kld:internal 0.0089; kld:public 0.0085; ppl:internal 1.5999; ppl:public 1.3466 | vs BF16 | L85 @ main |
| kacaforyah/Opti-27B earlier release of this publisher's file [L97,L103]coding-agent study [L114-115]publisher comparison build | Opti 27B (first release) | other | — | — | — | — | — | MMLU-Pro:scorerA 54.6; MMLU-Pro:scorerB 54.2 +/- 0.5 | other or no reference | L97 @ main | |
| kaushikvira/nvfp4full-dflash2-NInfer-v3 different artifact in the same family; cross-card rows measured in different sessions [L230-231]publisher comparison build | swift-abliterated v3 | NVFP4 | — | — | — | — | — | GSM8K-200 95.5% | reference not stated | L237 @ main | |
| kaushikvira/nvfp4full-dflash2-NInfer-v3 different artifact in the same family; cross-card rows measured in different sessions [L230-231]publisher comparison build | swift15 v3 | NVFP4 | — | — | — | — | — | GSM8K-200 95.0% | reference not stated | L238 @ main | |
| kaushikvira/nvfp4full-dflash2-NInfer-v3 different artifact in the same family; cross-card rows measured in different sessions [L230-231]publisher comparison build | thinkingcap v3 | NVFP4 | — | — | — | — | — | GSM8K-200 95.5% | reference not stated | L239 @ main | |
| kaushikvira/nvfp4full-dflash2-NInfer-v3 different artifact in the same family; cross-card rows measured in different sessions [L230-231]publisher comparison build | swift15-uncensored-ajgazin v3 | NVFP4 | — | — | — | — | — | GSM8K-200 94.5% | reference not stated | L240 @ main | |
| kaushikvira/swift15-nvfp4full-dflash2-NInfer-v3 Weights in VRAM [L126]publisher comparison build | Swift-1.0-abliterated (prev. prod) | NVFP4 | 19.33 | GPU weights | — | — | — | — | IFBench prompt-loose 70.3; IFBench instr-strict 67.7; IFBench instr-loose 71.2; GSM8K-200 95.5% (191/200) | other or no reference | L118 @ main |
| kaushikvira/swift15-nvfp4full-dflash2-NInfer-v3 Cross-session row, asterisked [L182]publisher comparison build | nvfp4full-dflash2.v3 | NVFP4 | — | — | — | — | — | GSM8K-200 96.5%* | other or no reference | L182 @ main | |
| kaushikvira/swift15-nvfp4full-dflash2-NInfer-v3 Cross-session row [L183]publisher comparison build | swift-abliterated v3 | NVFP4 | — | — | — | — | — | GSM8K-200 95.5% | other or no reference | L183 @ main | |
| kaushikvira/swift15-nvfp4full-dflash2-NInfer-v3 [L185]publisher comparison build | thinkingcap v3 | NVFP4 | — | — | — | — | — | GSM8K-200 95.5% | other or no reference | L185 @ main | |
| kaushikvira/swift15-nvfp4full-dflash2-NInfer-v3 [L186]publisher comparison build | swift15-uncensored-ajgazin v3 | NVFP4 | — | — | — | — | — | GSM8K-200 94.5% | other or no reference | L186 @ main | |
| kaushikvira/swift15-uncensored-ajgazin-nvfp4full-dflash2-NInfer-v3 comparison build [L69-70]publisher comparison build | Swift-1.5 (prod, censored) | NVFP4 | — | — | — | — | — | IFBench prompt-strict 69.0; GSM8K-200 95.0% | other or no reference | L69 @ main | |
| kaushikvira/swift15-uncensored-ajgazin-nvfp4full-dflash2-NInfer-v3 family table, asterisked [L184]; different sessions [L179]publisher comparison build | nvfp4full-dflash2.v3 | NVFP4 | — | — | — | — | — | IFBench prompt-strict 65.0*; GSM8K-200 96.5%* | other or no reference | L184 @ main | |
| kaushikvira/swift15-uncensored-ajgazin-nvfp4full-dflash2-NInfer-v3 family table [L185]publisher comparison build | swift-abliterated v3 | NVFP4 | — | — | — | — | — | IFBench prompt-strict 66.3; GSM8K-200 95.5% | other or no reference | L185 @ main | |
| kaushikvira/swift15-uncensored-ajgazin-nvfp4full-dflash2-NInfer-v3 family table [L187]publisher comparison build | thinkingcap v3 | NVFP4 | — | — | — | — | — | IFBench prompt-strict 68.0; GSM8K-200 95.5% | other or no reference | L187 @ main | |
| kaushikvira/thinkingcap-nvfp4full-dflash2-NInfer-v3 comparison build in A/B table [L137-141]publisher comparison build | Swift-1.5 (our prod) | NVFP4 | — | — | — | — | — | GSM8K-200 95.0% (190/200); IFBench prompt-strict 69.0; IFBench prompt-loose 72.7; IFBench instr-strict 70.4; IFBench instr-loose 73.6 | other or no reference | L137 @ main | |
| kaushikvira/thinkingcap-nvfp4full-dflash2-NInfer-v3 asterisked values in family table [L203]; cross-card rows from different sessions [L198]publisher comparison build | nvfp4full-dflash2.v3 | NVFP4 | — | — | — | — | — | GSM8K-200 96.5%*; IFBench prompt-strict 65.0* | other or no reference | L203 @ main | |
| kaushikvira/thinkingcap-nvfp4full-dflash2-NInfer-v3 family table [L204]; different sessions [L198]publisher comparison build | swift-abliterated v3 | NVFP4 | — | — | — | — | — | GSM8K-200 95.5%; IFBench prompt-strict 66.3 | other or no reference | L204 @ main | |
| kaushikvira/thinkingcap-nvfp4full-dflash2-NInfer-v3 family table [L207]; different sessions [L198]publisher comparison build | swift15-uncensored-ajgazin v3 | NVFP4 | — | — | — | — | — | GSM8K-200 94.5%; IFBench prompt-strict 70.3 | other or no reference | L207 @ main | |
| kayceesamuel/NF4DQ referencepublisher comparison build | BF16 | bitsandbytes | 54.65 | download | — | — | 6.9530 ppl:wikitext | — | — | vs BF16 | L24 @ main |
| kayceesamuel/NF4DQ throughput table lists 17.66 GiB for same build [L37]publisher comparison build | Q4_K_M | bitsandbytes | 16.81 | download | — | — | 6.9787 ppl:wikitext | — | — | vs BF16 | L25 @ main |
| kayceesamuel/NF4DQ publisher comparison build | Q3_K_M | bitsandbytes | 13.58 | download | — | — | 7.3362 ppl:wikitext | — | — | vs BF16 | L27 @ main |
| kdqemre/RQ2_K_L-GGUF publisher comparison build | Q5_K_M (Baseline) | GGUF | 19.50 | download | — | — | — | — | GSM8K-100 98 | other or no reference | L91 @ main |
| kingjones777/ROCmFP4-STRIX-MTP-GGUF quoted by singulared, not measuredother publisher's 40-chunk run, not comparable to 145-chunk [L229-233] | STRIX (40-chunk ppl) | GGUF | — | — | — | 5.8877 | — | — | other or no reference | L229 @ main | |
| kingjones777/ROCmFP4-STRIX-MTP-GGUF quoted by singulared, not measuredother publisher's 40-chunk run [L229] | FAST (40-chunk ppl) | GGUF | — | — | — | 5.9233 | — | — | other or no reference | L229 @ main | |
| lackonendes/PAW-27B-X3-GGUF IFBench cell is '31.2%<br>—' so c20 is loose only, strict absentMBPP cells for this row not in candidatespublisher comparison build | Q8 parent | GGUF | — | — | — | — | — | MMLU-Pro 65.2%; IFBench_loose 31.2%<br>—; HumanEval 97.0; HumanEval+ 91.5; GSM8K 94.0% | other or no reference | L90 @ main | |
| Launch80/PARO-int5 GSM8K is a range 97.4-97.6%; only the low end has an id [L81]no disk size in tablepublisher comparison build | PARO-MXFP4 | other | — | — | 90.3% | — | — | kld:top256 0.0419 | vs BF16 | L79 @ main | |
| libertyhhn/GPTQ-W4A16-g128 Dense bf16 size 55.6 GB from prose [L24]publisher comparison build | Dense (bf16) | GPTQ | 55.60 | download | — | — | — | — | ppl:wikitext2 6.16; MMLU 0-shot 84.89; MMLU 5-shot 84.21; GSM8K 97.35 | vs BF16 | L24 @ main |
| Litwein/oQ4e-DWQ-MTP-Vision-MLX Measured on a 16-batch subset [L134]publisher comparison build | Raw oQ4e-XL | MLX | — | — | — | — | — | kld:heldout_teacher 0.0579 | other or no reference | L129 @ main | |
| Litwein/oQ4e-DWQ-MTP-Vision-MLX publisher comparison build | M4Q init | MLX | — | — | — | — | — | kld:heldout_teacher 0.0287 | other or no reference | L129 @ main | |
| Litwein/oQ4e-DWQ-MTP-Vision-MLX KL 0.019 vs the bf16 base [L101-102]; different reference from the held-out teacher KL rowspublisher comparison build | oQ8e teacher | MLX | — | KL 0.019 vs | — | — | — | — | other or no reference | L101 @ main | |
| lmcoleman/MagicQuant-GGUF Searched and measured but not published; dominated by Q4 [L54-60]Cell holds ppl and percent increase vs BF16 together [L60]publisher comparison build | Q5 (withheld, not published) | GGUF | 18.98 | download | — | — | — | *6.7666 (+0.33%)* | ppl:wikitext-2 *6.7666 (+0.33%)* | vs BF16 | L60 @ main |
| lmcoleman/ROCmFPX-GGUF sibling repo not named [L33-34]publisher comparison build | sibling repo Q4_K_M (MagicQuant source) | GGUF | — | — | — | 6.7522 ppl:wikitext | — | — | vs BF16 | L33 @ main | |
| lmstudio-community/MLX-4bit measured by Steven10429made from bf16 [L28] | lmstudio-community/Qwen3.8-27B-MLX-4bit | MLX | 16.05 | download | — | — | — | — | ppl:wikitext2 6.4349 | vs BF16 | L32 @ main |
| lmstudio-community (repo not named on card) measured by AtomicChat[L215] | Q4_K_M (lmstudio-community) | 16.80 | download | 0.02094 | — | — | — | — | vs BF16 | L215 @ main | |
| Lottolabs/TT-Mixed-BFP4-BFP8-P150 Reference is the TT all-BFP8 build, not BF16 [L48]top-1 values are fractions as printed [L52-55]publisher comparison build | Round-to-nearest BFP4 (release dual-noc) vs TT all-BFP8 | NVFP4 | — | — | — | — | — | kld:chat 0.180; top1:chat 0.838; kld:code 0.053; top1:code 0.940; kld:instruct 0.013; top1:instruct 0.960; kld:all 0.103; top1:all 0.899 | vs BF16 | L52 @ main | |
| lribeiro/FP8-Pessoa [L250]publisher comparison build | FP8 W8A8 (GPTQ) | FP8 | — | 0.0102 | — | — | — | — | other or no reference | L250 @ main | |
| lribeiro/FP8-Pessoa [L251]publisher comparison build | FP8 W8A16 (ModelOpt) | FP8 | — | 0.0127 | — | — | — | — | other or no reference | L251 @ main | |
| lribeiro/FP8-Pessoa Labeled way too big [L275]publisher comparison build | gptq-fp8-att (best quality, any size) | FP8 | — | 0.0026 | — | — | — | — | other or no reference | L275 @ main | |
| lribeiro/FP8-Pessoa ~28 GB [L277]publisher comparison build | gptq-mxfp8-mixed | FP8 | — | 0.0107 | — | — | — | — | other or no reference | L277 @ main | |
| lribeiro/FP8-Pessoa ~27 GB [L278]publisher comparison build | gptq-nvfp4-mixed-8 | FP8 | — | 0.0109 | — | — | — | — | other or no reference | L278 @ main | |
| lribeiro/FP8-Pessoa [L279]publisher comparison build | v17 (best all-FP8 W8A8) | FP8 | — | 0.0123 | — | — | — | — | other or no reference | L279 @ main | |
| magiccodingman/MagicQuant-GGUF Struck through in table (replaced by MQ-Q6_K_1) [L35]publisher comparison build | LM-Q8_0 | GGUF | 29.05 | download | ~~0.000712~~ | — | — | — | — | reference not stated | L35 @ main |
| maglun/MLX-Mixed-3.80bpw measured on a different machine than the rest of the table [L131]publisher comparison build | BF16 text source (cross-host) | MLX | 55.57 | text weights (excl. vision, MTP) | — | — | — | — | ppl:heldout-4doc 5.8107; ppl:heldout-6doc 5.6728; HumanEval+ 75.0%; GSM8K 91.7% | vs BF16 | L74 @ main |
| maglun/MLX-Mixed-3.80bpw publisher's own earlier release [L98]publisher comparison build | mixed affine 4.95 bpw (sibling release) | MLX | 17.57 | download | — | — | — | — | ppl:heldout-4doc 5.9056; ppl:heldout-6doc 5.7428; HumanEval+ 75.0%; GSM8K 100.0% | vs BF16 | L74 @ main |
| maglun/MLX-Mixed-3.80bpw builder of this MXFP4 file not stated on card [L100]publisher comparison build | MXFP4, with BF16 vision + MTP drafter | MLX | 15.21 | download | — | — | — | — | HumanEval+ 75.0%; GSM8K 100.0% | vs BF16 | L100 @ main |
| maglun/MLX-Mixed-3.80bpw builder of this MXFP4 file not stated on card [L99]publisher comparison build | MXFP4, text only | MLX | 14.29 | text weights (excl. vision, MTP) | — | — | — | — | HumanEval+ 75.0%; GSM8K 100.0% | vs BF16 | L99 @ main |
| maglun/MLX-Mixed-4.95bpw measured on a different machine, not host-matched [L127]publisher comparison build | BF16 text source (cross-host) | MLX | 55.57 | download | — | — | 5.8107 | — | HumanEval+ 75.0%; GSM8K 91.7% | vs BF16 | L93 @ main |
| maglun/MLX-Mixed-4.95bpw author's sibling release [L97]publisher comparison build | sibling release - mixed affine 3.80 bpw | MLX | 13.70 | download | — | — | 6.4916 | — | HumanEval+ 75.0%; GSM8K 91.7% | vs BF16 | L97 @ main |
| malaiwah/EXL3-EDA-research incumbent hand-designed recipe [L93]re-captured row [L94] is bit-identical and omitted as duplicatepublisher comparison build | hydrated, published | EXL3 | — | 0.002699883159684943 | 97.797 % | — | — | — | other or no reference | L93 @ main | |
| malaiwah/EXL3-K5K6 resident 17.89 GiB; size cell asterisked [L196]publisher comparison build | K4 (v5 10M) | EXL3 | 28.31 | download | 0.010604 | 95.76 % | — | — | — | vs BF16 | L196 @ main |
| malaiwah/EXL3-K5K6 resident 20.31 GiB [L193][L287]publisher comparison build | K5K6-hydrated (v5 10M) | EXL3 | 21.61 | download | 0.002760 | 97.70 % | — | — | — | vs BF16 | L193 @ main |
| malaiwah/EXL3-K5K6 shard-0, 512 contexts [L455]; payload bytespublisher comparison build | K5K6-hydrated (shard-0 replay) | EXL3 | 21.60 | download | 0.002700 | 97.80 % | — | — | — | vs BF16 | L455 @ main |
| malaiwah/EXL3-K5K6 resident 18.41 GiB; v5 value excludes int8 input overlay [L40,L195]publisher comparison build | K5K6-context (v5 10M) | EXL3 | 20.70 | download | 0.003509 | 97.44 % | — | — | — | vs BF16 | L195 @ main |
| malaiwah/EXL3-K5K6 shard-0 [L457]; payload bytespublisher comparison build | K5K6-context (shard-0 replay) | EXL3 | 20.69 | download | 0.003409 | 97.55 % | — | — | — | vs BF16 | L457 @ main |
| malaiwah/EXL3-K5K6 shard-0 [L461]publisher comparison build | K4 (shard-0 replay) | EXL3 | — | 0.010345 | 95.91 % | — | — | — | vs BF16 | L461 @ main | |
| malaiwah/EXL3-K5K6-context resident c7 20.32 GiB [L138]MMLU-Pro 55/70 [L746]publisher comparison build | K5K6 (online K5/K6) | EXL3 | 30.60 | download | 0.003210 | 97.52 % | — | — | kld:v3_corrected 0.007945; kld:v4 0.003455; top1:v4 97.50 %; MMLU-Pro 55/70 (78.6 %); kld:shard0 0.003141; top1:shard0 97.61 % | vs BF16 | L138 @ main |
| malaiwah/EXL3-K5K6-context resident c18 17.89 GiB [L140]asterisk after download size on card [L140]MMLU-Pro 57/70 [L743]publisher comparison build | K4 | EXL3 | 28.31 | download | 0.010604 | 95.76 % | — | — | kld:v3_corrected 0.029679; MMLU-Pro 57/70 (81.4 %); kld:shard0 0.010345; top1:shard0 95.91 % | vs BF16 | L140 @ main |
| malaiwah/EXL3-K5K6-context resident c2 20.31 GiB [L137]MMLU-Pro 56/70 [L744]publisher comparison build | K5K6-hydrated | EXL3 | 21.61 | download | 0.002760 | 97.70 % | — | — | kld:v3_corrected 0.007172; kld:v4 0.003093; top1:v4 97.63 %; MMLU-Pro 56/70 (80.0 %); kld:shard0 0.002700; top1:shard0 97.80 % | vs BF16 | L137 @ main |
| malaiwah/EXL3-K5K6-hydrated [L321]publisher comparison build | online K5/K6 sibling v5 10M | EXL3 | 21.82 | resident | 0.003210 | 97.52 % | — | — | — | other or no reference | L321 @ main |
| malaiwah/EXL3-K5K6-hydrated [L780]publisher comparison build | BF16-attention sibling v3 136-context | EXL3 | 21.82 | resident | 0.008157 | 96.97 % | — | — | — | other or no reference | L780 @ main |
| malaiwah/EXL3-K5K6-hydrated [L497] payload bytespublisher comparison build | context edition vLLM shard0 | EXL3 | 20.69 | download | 0.003409 | 97.55 % | — | — | — | other or no reference | L497 @ main |
| malaiwah/EXL3-K5K6-hydrated [L322] excludes int8 input overlay [L40]publisher comparison build | context edition v5 10M | EXL3 | 19.77 | resident | 0.003509 | 97.44 % | — | — | — | other or no reference | L322 @ main |
| malaiwah/EXL3-K5K6-hydrated [L324]publisher comparison build | K4 v5 10M | EXL3 | 19.21 | resident | 0.010604 | 95.76 % | — | — | — | other or no reference | L324 @ main |
| malaiwah/EXL3-K5K6-hydrated [L496]publisher comparison build | online K5/K6 vLLM shard0 | EXL3 | — | 0.003141 | 97.61 % | — | — | — | other or no reference | L496 @ main | |
| malaiwah/EXL3-K5K6-hydrated [L501]publisher comparison build | K4 shard0 | EXL3 | — | 0.010345 | 95.91 % | — | — | — | other or no reference | L501 @ main | |
| malaiwah/EXL3-K5K6-hydrated [L854]publisher comparison build | K5/K6 online K6 v4 36-context | EXL3 | — | 0.003455 | 97.50 % | — | — | — | other or no reference | L854 @ main | |
| malaiwah/EXL3-K5K6-hydrated [L855]publisher comparison build | context edition v4 36-context | EXL3 | — | 0.003990 | 97.36 % | — | — | — | other or no reference | L855 @ main | |
| malaiwah/EXL3-K6-parity same engine and reference as this build [L113]publisher comparison build | hydrated K5/K6 | EXL3 | 21.59 | download | 0.002700 | 97.797 % | — | — | — | vs BF16 | L110 @ main |
| malaiwah/EXL3-S16-V-research nearest neighbour build [L74]percentile columns marked * are binned approximations [L131,L135]publisher comparison build | K4 | EXL3 | — | 0.010345 | 95.91 % | — | — | kld_99 0.119353* | vs BF16 | L131 @ main | |
| malaiwah/EXL3-S16-V-research recommended build by same publisher [L37]percentile columns marked * are binned approximations [L130]publisher comparison build | context edition (malaiwah/Qwen3.8-27B-EXL3-K5K6-context) | EXL3 | — | 0.003409 | 97.55 % | — | — | kld_99 0.035699* | vs BF16 | L130 @ main | |
| malekoo/MLX-4bit ~54 GB bf16 MLX artifact [L22]publisher comparison build | bf16 MLX conversion | MLX | 54.00 | download | — | — | — | — | ppl:wikitext2 6.9352 | vs BF16 | L22 @ main |
| malekoo/MLX-8bit reference build, ~54 GB on disk [L22]publisher comparison build | bf16 MLX conversion | MLX | 54.00 | download | — | — | — | — | ppl:wikitext-2 6.9352 | vs BF16 | L22 @ main |
| malekoo/MLX-8bit 4-bit perplexity only in prose [L33]publisher comparison build | MLX 4-bit | MLX | — | — | — | — | — | ppl:wikitext-2 7.0871 | vs BF16 | L33 @ main | |
| malekoo/MLX-bf16 measured in identical harness [L126-129]publisher comparison build | malekoo/Qwen3.8-27B-MLX-4bit | MLX | 15.00 | unknown | — | — | 7.0871 | — | — | other or no reference | L129 @ main |
| manitcor/Obliterated-E03-bnb-4bit parent checkpointpublisher comparison build | BF16 E03 | bitsandbytes | — | — | — | 3.2099 | — | capability checks 83.3% | vs BF16 | L140 @ main | |
| MESHIVEAI/Qwen-3.8-27B-Q4_K_M-Imatrix same F16 GGUF and Q4_K_M recipe, same tensor size [L47-L49]publisher comparison build | Q4_K_M, no imatrix | other | — | 0.025196 ± 0.000291 | 92.97% ± 0.07 | 6.9755 | — | ppl:korean-wikipedia 6.9596; ppl_delta_pct:korean-wikipedia +1.18% | other or no reference | L54 @ main | |
| MESHIVEAI/Qwen-3.8-27B-Q4_K_M-Imatrix unquantized baseline [L61, L79]publisher comparison build | F16 reference | other | — | — | — | — | — | ppl:korean-wikipedia 6.8782 | other or no reference | L61 @ main | |
| Mihai-LeanZero/Atlassian-Q6_K-GGUF Q8_0 file measured mean KLD 0.000842 under same procedure [L33]publisher comparison build | Q8_0 | GGUF | — | 0.000842 | — | — | — | — | vs BF16 | L33 @ main | |
| Mihai-LeanZero/Atlassian-Q8_0-GGUF mean KLD 0.00306 vs same bf16 GGUF [L33]publisher comparison build | Q6_K | GGUF | — | 0.00306 | — | — | — | — | vs BF16 | L33 @ main | |
| Minachist/INT5-Flat-5.7bpw-AutoRound reference [L38-39]publisher comparison build | BF16 base | AutoRound | — | — | — | 7.9112 ppl:wikitext | — | top1:wikitext 100 % | vs BF16 | L38 @ main | |
| Minachist/INT6-Flat-6.6bpw-AutoRound reference PPL [L40]publisher comparison build | BF16 base | AutoRound | — | — | — | 7.9112 | — | — | vs BF16 | L40 @ main | |
| Minachist/INT6-Mixed-AutoRound reference row [L35-43]publisher comparison build | BF16 base | AutoRound | — | — | 100 % | 7.9112 | — | — | vs BF16 | L35 @ main | |
| minima-ai/mnma_qwen3.8_27b_nvfp4 BF16 baseline served identically [L69-70]Weights in VRAM 50.13 GiB [L81]publisher comparison build | BF16 | NVFP4 | 53.83 | GPU weights | — | — | — | — | ppl:wikitext2-4k 6.95; ppl:wikitext2-32k 10.35; MMLU-Pro 80.4; GSM8K 95.5; AIME25 86.7; GPQA-Diamond 86.5; LiveCodeBench-v6 79.0 | vs BF16 | L74 @ main |
| mlasli/Heretic-Abliterated-MLX-6bit GSM8K delta called a measurement artifact [L85]publisher comparison build | v2.1.0 BF16 (mlasli/Qwen3.8-27B-Heretic-Uncensored-BF16) | MLX | — | — | — | — | — | MMLU 82.62; ARC-Challenge 57.42; HellaSwag 82.48; Winogrande 75.69; GSM8K flexible-extract 58.38; GSM8K strict-match 58.91 | vs BF16 | L78 @ main | |
| mlx-community/4bit measured by abuarchiveOfficial 4-bit release measured by card author [L25, L29-31]Prompt battery completed 15 of 16 [L35] | 4-bit (mlx-community) | MLX | 15.00 | download | — | — | — | — | ppl:tulu-3-sft-mixture 5.405 ± 0.159 | other or no reference | L29 @ main |
| mlx-community/4bit measured by leonsarmientorun by card author on same harness [L56] | Qwen3.8-27B-4bit (vanilla uniform) | MLX | — | — | — | — | — | MMLU 78.0%; MMLU_PRO 56.0% | other or no reference | L88 @ main | |
| mlx-community/mxfp4 measured by avlp12[L374] | mxfp4 (mlx-community) | MLX | 14.21 | download | 0.1437 | 86.4% | — | — | — | vs BF16 | L374 @ main |
| mlx-community/nvfp4 measured by avlp12repo id inferred from L46 listing of mlx-community/Qwen3.8-27B-{4bit,8bit,mxfp4,nvfp4} [L373] | nvfp4 (mlx-community) | MLX | 15.01 | download | 0.0962 | 89.0% | — | — | — | vs BF16 | L373 @ main |
| neroued/NInfer measured by WarlaxZUpstream revision 1cbd84e7 measured by card author [L71, L80-88] | Upstream NInfer encoding and layout | other | — | 0.0376 | — | — | — | ppl:perplexity-1m 4.3428; ppl:heldout-2026-09 4.3396 | other or no reference | L80 @ main | |
| nom666/Qwopus3.8-27B-MTPLX-8bit-Quality Speed-focused sibling [L19]publisher comparison build | Qwopus3.8-27B-MTPLX-4bit-Speed | other | — | — | — | — | — | non-thinking suite 37/38 | other or no reference | L44 @ main | |
| not named on card measured by TechnoBaptistFP16 reference [L212] | Qwen3.8-27B FP16 | 54.00 | download | — | — | — | — | MMLU-Redux 91.46; MuSR 79.63; GSM8K 97.19; MATH-500 99.80; AIME25 96.67; AIME26 94.58; HumanEval+ 93.29; MBPP+ 83.86; LiveCodeBench 90.05; IFEval 91.50; IFBench 71.00; BFCL v3 76.74; MMMU-Pro 81.73; OCR Bench v2 60.99 | other or no reference | L212 @ main | |
| not named on card measured by drawthingsaiBF16 reference run by card author [L73] | Qwen3.8-27B-BF16 | 49.98 | download | — | — | — | — | AIME 2026 100.00% | vs BF16 | L73 @ main | |
| not named on card measured by TelperionAIsize ~30 GB [L32] | Qwen FP8 (reference, vLLM) | FP8 | 30.00 | download | — | 96.15% | — | — | — | vs BF16 | L32 @ main |
| not named on card measured by TelperionAIdisagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38] | Qwen3.8-27B-FP8 (official) | FP8 | 30.00 | download | — | 96.15% | — | — | — | vs BF16 | L47 @ main |
| not named on card measured by TelperionAIsize ~30 GB [L32] | Qwen FP8 (reference, vLLM) | FP8 | 30.00 | download | — | 96.15% | — | — | — | vs BF16 | L32 @ main |
| not named on card measured by TelperionAIdisagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38] | Qwen3.8-27B-FP8 (official) | FP8 | 30.00 | download | — | 96.15% | — | — | — | vs BF16 | L47 @ main |
| not named on card measured by TelperionAIsize ~30 GB [L32] | Qwen FP8 (reference, vLLM) | FP8 | 30.00 | download | — | 96.15% | — | — | — | vs BF16 | L32 @ main |
| not named on card measured by TelperionAIQwen's own FP8 release measured by card author [L14-15, L34] | Qwen FP8 (same format, no AWQ) | AWQ | 30.00 | unknown | — | 96.15% | — | — | — | vs BF16 | L34 @ main |
| not named on card measured by glydmeasured by card author, repo not named [L27] | Qwen FP8 | FP8 | 29.50 | download | 0.00465 | — | — | — | — | vs BF16 | L27 @ main |
| not named on card measured by glyd[L27] | Qwen FP8 | FP8 | 29.50 | download | 0.00465 | — | — | — | — | vs BF16 | L27 @ main |
| not named on card measured by tooltdprovider not given in row [L60] | UD-Q8_K_XL | 29.30 | download | 0.000850 | 98.970% | 6.953800 | — | — | vs BF16 | L60 @ main | |
| not named on card measured by malaiwahllama.cpp cross-engine; engine control 0.000507 mean not subtracted [L467-470] | GGUF Q8_0 llama.cpp | GGUF | 29.04 | download | 0.001087 | 98.53 % | — | — | — | vs BF16 | L452 @ main |
| not named on card measured by malaiwahshard-0 of own suite for c71/c72, cross-engine, not directly comparable with vLLM rows [L459-461,L619-621]wikitext columns are llama.cpp's own protocol, measured by card author [L647-660] | GGUF Q8_0 llama.cpp | GGUF | 29.04 | download | 0.001087 | 98.53 % | — | — | kld:wikitext 0.000926 ± 0.000042; top1:wikitext 98.761 % | vs BF16 | L461 @ main |
| not named on card measured by malaiwahdifferent engine, not directly comparable [L689] | GGUF Q8_0 llama.cpp shard0 | GGUF | 29.04 | download | 0.001087 | 98.53 % | — | — | — | other or no reference | L492 @ main |
| not named on card measured by Ar3ntizsize approximate (about 28.6) [L58] | MLX affine 8-bit | MLX | 28.60 | download | 0.002 | — | — | — | — | vs BF16 | L58 @ main |
| not named on card measured by Jianqiao1[L62] | Unsloth UD-Q6_K_XL | 25.92 | download | 0.001331 | 98.493% | 6.953912 | — | kld_99 0.011536 | vs BF16 | L62 @ main | |
| not named on card measured by TelperionAI[L36] | RadixArk NVFP4 | NVFP4 | 24.70 | unknown | — | 90.23% | — | — | — | vs BF16 | L36 @ main |
| not named on card measured by TelperionAI[L37] | unsloth NVFP4 | NVFP4 | 24.70 | unknown | — | 91.75% | — | — | — | vs BF16 | L37 @ main |
| not named on card measured by tooltd | UD-Q6_K_XL | 24.14 | download | 0.001380 | 98.520% | 6.953600 | — | — | vs BF16 | L61 @ main | |
| not named on card measured by malaiwah[L454] | GGUF Q6_K llama.cpp | GGUF | 22.88 | download | 0.002035 | 97.98 % | — | — | — | vs BF16 | L454 @ main |
| not named on card measured by malaiwahcross-engine [L463]wikitext protocol [L661] | GGUF Q6_K llama.cpp | GGUF | 22.88 | download | 0.002035 | 97.98 % | — | — | kld:wikitext 0.002286 ± 0.000108; top1:wikitext 97.875 % | vs BF16 | L463 @ main |
| not named on card measured by malaiwah[L494] | GGUF Q6_K llama.cpp shard0 | GGUF | 22.88 | download | 0.002035 | 97.98 % | — | — | — | other or no reference | L494 @ main |
| not named on card measured by malaiwahdifferent engine (llama.cpp) so not a format-only comparison [L25,L116-119] | GGUF Q6_K (llama.cpp) | GGUF | 22.88 | download | 0.002035 | 97.980 % | — | — | — | vs BF16 | L111 @ main |
| not named on card measured by tooltd | Q6_K | 21.31 | download | 0.002290 | 97.860% | 6.950700 | — | — | vs BF16 | L62 @ main | |
| not named on card measured by finalventpublicly distributed NVFP4 of the same base; repo not named [L95] | Generic 4-bit (public NVFP4) | NVFP4 | 20.60 | download | — | — | — | — | ReferenceFactsCaptured 0.53; ReferenceFactsContradictedPct 12.4%; OverconfidentPerPhoto 0.97; ClearErrorsPerPhoto 0.43 | vs BF16 | L77 @ main |
| not named on card measured by malaiwah[L459] | GGUF UD-Q5_K_XL llama.cpp | GGUF | 20.22 | download | 0.004444 | 97.20 % | — | — | — | vs BF16 | L459 @ main |
| not named on card measured by malaiwahcross-engine [L468]wikitext protocol [L662] | GGUF UD-Q5_K_XL llama.cpp | GGUF | 20.22 | download | 0.004444 | 97.20 % | — | — | kld:wikitext 0.004426 ± 0.000167; top1:wikitext 97.178 % | vs BF16 | L468 @ main |
| not named on card measured by malaiwah[L499] | GGUF UD-Q5_K_XL llama.cpp shard0 | GGUF | 20.22 | download | 0.004444 | 97.20 % | — | — | — | other or no reference | L499 @ main |
| not named on card measured by Jianqiao1[L61] | Unsloth UD-Q5_K_XL | 20.22 | download | 0.004526 | 97.153% | 6.967300 | — | kld_99 0.040594 | vs BF16 | L61 @ main | |
| not named on card measured by TelperionAI[L38] | Inferact NVFP4 | NVFP4 | 20.10 | unknown | — | 84.11% | — | — | — | vs BF16 | L38 @ main |
| not named on card measured by kingjones777measured by card author; ggml-org build, repo not named for this file [L92] | Q4_K_M (reference, ggml-org) | 18.97 | download | — | — | 5.8926 | — | — | other or no reference | L92 @ main | |
| not named on card measured by tooltd | Q5_K_M | 18.47 | download | 0.006220 | 96.700% | 6.974200 | — | — | vs BF16 | L63 @ main | |
| not named on card measured by Jianqiao1[L60] | Unsloth UD-Q4_K_XL | 17.92 | download | 0.008640 | 96.073% | 6.976017 | — | kld_99 0.090719 | vs BF16 | L60 @ main | |
| not named on card measured by trevestppl_delta_pct is vs Q8_0 [L27] | unsloth UD-Q4_K_XL | 17.92 | download | — | — | 7.0057 | +0.24% | — | vs a Q8_0 | L27 @ main | |
| not named on card measured by TechnoBaptistConventional 4-bit comparison build [L213] | Qwen3.8-27B UD-Q4_K_XL | 17.60 | download | — | — | — | — | MMLU-Redux 93.35; MuSR 73.01; GSM8K 96.66; MATH-500 99.40; AIME25 92.91; AIME26 93.00; HumanEval+ 95.73; LiveCodeBench 87.96; IFEval 88.83; IFBench 65.65; BFCL v3 75.05; OCR Bench v2 65.45 | other or no reference | L213 @ main | |
| not named on card measured by varshavvvvvMBPP+ and MMMU-Pro cells blank [L249,L254] | Qwen3.8-27B UD-Q4_K_XL | 17.60 | text weights (excl. vision, MTP) | — | — | — | — | MMLU-Redux 93.35; MuSR 73.01; GSM8K 96.66; MATH-500 99.40; AIME25 92.91; AIME26 93.00; HumanEval+ 95.73; LiveCodeBench 87.96; IFEval 88.83; BFCL v3 75.05; OCR Bench v2 65.45 | other or no reference | L213 @ main | |
| not named on card measured by kacaforyahmedian KLD 0.0065 also in table [L77] | llama.cpp Q4_K_M | 16.80 | download | 0.0156 | 94.7% | 6.457 | 1.000x | ppl_ratio:paired 1.0031 +/- 0.0011; kld_99 0.146; ppl_delta_pct:wikitext-2 +0.31% +/- 0.11%; ppl_delta_pct:C4 +1.0% +/- 0.2%; ppl_delta_pct:python-stdlib +3.1% +/- 0.5%; ppl_delta_pct:pride-and-prejudice +7.9% +/- 0.8%; MMLU-Pro:scorerB 57.14 +/- 0.45 | other or no reference | L35 @ main | |
| not named on card measured by andrew-paul[L31] | Stock Q4_K_M | 16.73 | download | 0.016437 | — | — | — | — | vs BF16 | L31 @ main | |
| not named on card measured by tooltd | UD-Q4_K_XL | 16.69 | download | 0.008606 | 96.091% | 6.979220 | — | — | vs BF16 | L64 @ main | |
| not named on card measured by singularedPPL +/- 0.0402; -0.033 vs STRIX (c46) [L144] | mainline Q4_K_M (reference) | 16.55 | download | — | — | 6.3383 ± 0.0402 | — | — | other or no reference | L144 @ main | |
| not named on card measured by tooltd | Q4_K_M | 15.93 | download | 0.015490 | 94.650% | 6.956100 | — | — | vs BF16 | L69 @ main | |
| not named on card measured by tooltd | Autoround-Q4_K_M (Autoround) | AutoRound | 15.66 | download | 0.014657 | 94.859% | 6.950294 | — | — | vs BF16 | L67 @ main |
| not named on card measured by Ar3ntizsize approximate (about 15.2) [L57] | MLX affine 4-bit | MLX | 15.20 | download | 0.045 | — | — | — | — | vs BF16 | L57 @ main |
| not named on card measured by tooltdEvalPlus row named Qwen3.8-27B-IQ4_NL [L38] | IQ4_NL | 15.20 | download | 0.018427 | 94.230% | 7.006472 | — | HumanEval 0.951; HumanEval+ 0.915; MBPP 0.902; MBPP+ 0.778 | vs BF16 | L38 @ main | |
| not named on card measured by tooltd | Q4_K_S | 15.01 | download | 0.018921 | 94.235% | 6.966826 | — | — | vs BF16 | L76 @ main | |
| not named on card measured by avlp12[L372] | orcarouter 4-bit | 14.95 | download | 0.0754 | 90.3% | — | — | — | vs BF16 | L372 @ main | |
| not named on card measured by tooltd | Q4_K_S-i1 | 14.74 | download | 0.019805 | 93.996% | 6.989551 | — | — | vs BF16 | L78 @ main | |
| not named on card measured by tooltd | Q4_0-AutoRound-Code (webhie) | AutoRound | 14.64 | download | 0.026586 | 92.970% | 7.067142 | — | — | vs BF16 | L81 @ main |
| not named on card measured by tooltd | IQ4_XS | 14.63 | download | 0.018652 | 94.270% | 7.012695 | — | — | vs BF16 | L73 @ main | |
| not named on card measured by tooltd | UD3-Q4_K_S (unsloth UD3) | 14.30 | download | 0.013652 | 95.149% | 6.969514 | — | — | vs BF16 | L66 @ main | |
| not named on card measured by tooltd | IQ4_XS-i1 | 14.26 | download | 0.019271 | 94.141% | 7.012810 | — | — | vs BF16 | L77 @ main | |
| not named on card measured by kacaforyahKLD row measured on L40S with re-made file [L81] | llama.cpp Q3_K_M | 13.50 | download | 0.0496 | 90.4% | 6.708 | 1.039x | ppl_ratio:paired 1.0395 +/- 0.0021; kld_99 0.504; MMLU-Pro:scorerB 56.50 +/- 0.45 | other or no reference | L36 @ main | |
| not named on card measured by Jianqiao1[L59] | Unsloth UD-Q3_K_XL | 13.44 | download | 0.031355 | 92.396% | 7.111811 | — | kld_99 0.353498 | vs BF16 | L59 @ main | |
| not named on card measured by tooltd | UD3-IQ4_XS (unsloth UD3) | 13.27 | download | 0.018772 | 93.975% | 7.004732 | — | HumanEval 0.890; HumanEval+ 0.866; MBPP 0.897; MBPP+ 0.751 | vs BF16 | L46 @ main | |
| not named on card measured by tooltd | IQ4_XS-Smaller_3.96 (jrell) | 12.61 | download | 0.055499 | 90.090% | 7.252766 | — | — | vs BF16 | L91 @ main | |
| not named on card measured by Bucoidcomparison build of similar size, measured by this card's author [L11]kld_99 is the 99.9% KL quantile [L25] | UD_IQ3_K_XL | 12.50 | download | 0.03130 ± 0.00032 | 92.419% ± 0.069% | 7.1117 ± 0.0459 | — | — | vs BF16 | L19 @ main | |
| not named on card measured by Bucoidbaseline same-size build; publisher/repo not named; card calls it roughly 3-bit [L61,L54] | UD_IQ3_K_XL | 12.50 | download | 0.03130 ± 0.00032 | 92.419% ± 0.069% | 7.1117 ± 0.0459 | — | kld_max 21.409; kld_999 1.219 | other or no reference | L60 @ main | |
| not named on card measured by andrew-paul[L32] | Stock Q3_K_S | 12.26 | download | 0.088238 | — | — | — | — | vs BF16 | L32 @ main | |
| not named on card measured by tooltd | IQ4_XS-3.84bpw | 12.18 | download | 0.049513 | 90.989% | 7.158078 | — | — | vs BF16 | L86 @ main | |
| not named on card measured by Jianqiao1[L58] | Unsloth UD-IQ3_XXS | 11.91 | download | 0.054517 | 90.359% | 7.246240 | — | kld_99 0.627471 | vs BF16 | L58 @ main | |
| not named on card measured by kacaforyahincludes unused MTP block 0.45 GB [L47] | llama.cpp IQ3_XXS | 11.60 | download | — | — | 6.928 | 1.073x | MMLU-Pro:scorerA 53.3 | other or no reference | L38 @ main | |
| not named on card measured by talxcc[L64] | Standard Q3_K_L | 11.40 | download | — | — | — | — | Coding_Retention ~94.2% | other or no reference | L64 @ main | |
| not named on card measured by tooltd | GSQ-RCO-IQ3_S (ISTA) | 11.29 | download | 0.055475 | 89.657% | 7.062697 | — | — | vs BF16 | L89 @ main | |
| not named on card measured by sdkyuan[L72-76] | Unsloth UD-IQ3_XXS | 10.93 | download | — | — | — | — | top1:reasoning 95.7%; kld:reasoning 0.018; top1:code 94.6%; kld:code 0.038; top1:tool_calling 84.0%; kld:tool_calling 0.929; top1:wikitext 90.0%; kld:wikitext 0.064 | other or no reference | L72 @ main | |
| not named on card measured by Jianqiao1[L57] | Unsloth UD-Q2_K_XL | 10.68 | download | 0.091579 | 86.160% | 7.403969 | — | kld_99 0.925859 | vs BF16 | L57 @ main | |
| not named on card measured by kacaforyah[L114-115] | llama.cpp IQ2_M | 10.50 | download | — | — | 7.318 | 1.134x | MMLU-Pro:scorerA 53.2 | other or no reference | L39 @ main | |
| not named on card measured by Jianqiao1[L56] | Unsloth UD-IQ2_M | 10.32 | download | 0.103564 | 85.496% | 7.507406 | — | kld_99 1.076759 | vs BF16 | L56 @ main | |
| not named on card measured by andrew-paulKLD printed as approximate (about 0.1551) [L33] | Stock IQ2_M | 10.19 | download | ≈0.1551 | — | — | — | — | vs BF16 | L33 @ main | |
| not named on card measured by wiklif[L118] | ISTA-DASLab GSQ-RCO IQ3_XXS | 10.09 | download | 0.0783 kld:prose | — | — | — | kld:code 0.0602 | vs a Q8_0 | L118 @ main | |
| not named on card measured by drawthingsaiCommunity baseline, same profile [L76,L173] | AD-IQ2_XS | 9.9 | download | — | — | — | — | AIME 2026 96.67%; GPQA Diamond 76.77%; IFEval 90.20%; BFCL v4 (1k) 68.46% | vs BF16 | L76 @ main | |
| not named on card measured by sdkyuan[L56-60] | Unsloth UD-Q2_K_XL | 9.8 | download | — | — | — | — | top1:reasoning 94.4%; kld:reasoning 0.032; top1:code 93.1%; kld:code 0.061; top1:tool_calling 82.5%; kld:tool_calling 1.026; top1:wikitext 87.5%; kld:wikitext 0.098 | other or no reference | L56 @ main | |
| not named on card measured by drawthingsaiCommunity baseline, same profile [L74,L173] | UD-Q2_K_XL | 9.8 | download | — | — | — | — | AIME 2026 100.00%; GPQA Diamond 85.35%; IFEval 93.89%; BFCL v4 (1k) 69.46% | vs BF16 | L74 @ main | |
| not named on card measured by TechnoBaptistConventional '2-bit' comparison build, real 2.8 bpw [L87,L214] | Qwen3.8-27B IQ2_XXS | 9.4 | download | — | — | — | — | MMLU-Redux 88.93; MuSR 66.99; GSM8K 89.90; MATH-500 84.60; AIME25 66.67; AIME26 57.50; HumanEval+ 91.46; MBPP+ 78.89; LiveCodeBench 56.40; IFEval 84.03; IFBench 53.76; BFCL v3 70.28; MMMU-Pro 65.19; OCR Bench v2 61.70 | other or no reference | L214 @ main | |
| not named on card measured by varshavvvvvdescribed as a widely-used 2-bit build [L87] | Qwen3.8-27B IQ2_XXS | 9.4 | text weights (excl. vision, MTP) | — | — | — | — | MMLU-Redux 88.93; MuSR 66.99; GSM8K 89.90; MATH-500 84.60; AIME25 66.67; AIME26 57.50; HumanEval+ 91.46; MBPP+ 78.89; LiveCodeBench 56.40; IFEval 84.03; BFCL v3 70.28; MMMU-Pro 65.19; OCR Bench v2 61.70 | other or no reference | L214 @ main | |
| not named on card measured by Jianqiao1[L55] | Unsloth UD-IQ2_XXS | 9 | download | 0.163689 | 82.389% | 7.957195 | — | kld_99 1.739405 | vs BF16 | L55 @ main | |
| not named on card measured by lackonendes[L43] | AtomicChat AD-IQ2XXS | 9 | download | — | — | — | — | MMLU-Pro 53.0%; GSM8K 94%; HumanEval+ 73.8%; MBPP+ 71.7% | other or no reference | L43 @ main | |
| not named on card measured by drawthingsaiCommunity baseline, same profile [L77,L173] | AD-IQ1_M | 8.5 | download | — | — | — | — | AIME 2026 90.00%; GPQA Diamond 51.01%; IFEval 78.37%; BFCL v4 (1k) 62.08% | vs BF16 | L77 @ main | |
| not named on card measured by drawthingsaiCommunity baseline, same profile [L75,L173] | UD-IQ2_S | 8.4 | download | — | — | — | — | AIME 2026 96.67%; GPQA Diamond 79.29%; IFEval 90.39%; BFCL v4 (1k) 68.26% | vs BF16 | L75 @ main | |
| not named on card measured by sdkyuan[L64-68] | Unsloth UD-IQ2_S | 8.4 | download | — | — | — | — | top1:reasoning 92.9%; kld:reasoning 0.052; top1:code 91.5%; kld:code 0.089; top1:tool_calling 78.9%; kld:tool_calling 1.271; top1:wikitext 84.5%; kld:wikitext 0.148 | other or no reference | L64 @ main | |
| not named on card measured by anemlldecoded to BF16, scored in PyTorch [L87]; bpw in size cell | Mirai S (2.4-bit trellis) | 8.1 | text weights (excl. vision, MTP) | 0.1279 kld:heldout | 85.1% | — | — | kld:dev 0.1226; kld:dev_full 0.2716; kld_99 5.73; kld:wikitext 0.1674 | vs BF16 | L94 @ main | |
| not named on card measured by anemlldecoded to BF16, scored in PyTorch [L80] | Mirai S (2.4-bit trellis) | 8.1 | text weights (excl. vision, MTP) | 0.1279 kld:heldout | 85.1% | — | — | kld:dev 0.1226; kld:dev_full 0.2716; kld_99 5.73; kld:wikitext 0.1674 | vs BF16 | L87 @ main | |
| not named on card measured by oshitekuFootnote 1 only: 7.3 GB, 84.0% / 0.311 [L77] | UD-IQ2_XXS | 7.3 | download | 0.311 | 84.0 | — | — | — | vs BF16 | L77 @ main | |
| not named on card measured by lackonendes[L42] | Unsloth IQ2XXS | 7.3 | download | — | — | — | — | MMLU-Pro 46.2%; GSM8K 93%; HumanEval+ 84.1%; MBPP+ 72.8% | other or no reference | L42 @ main | |
| not named on card measured by CodeMasterCody3Dmeasured head-to-head on same binary, kernels and text per card [L102]; row label reads 'wikitext c512 PPL 11.8346 (with Doctors)' [L108]card states 2.125 bpw for Bonsai in the size cell [L107] | Ternary-Bonsai-27B (with Doctors comparison row) | 7.2 | download | — | — | 11.01 ppl:wikitext | — | — | reference not stated | L106 @ main | |
| not named on card measured by Ar3ntizcolumn alignment of row L65 uncertain (BBH/Long context labels) [L62,L65] | GGUF Q8_0 (llama.cpp) | GGUF | — | — | — | — | — | GSM8K 18; MATH 13 | vs BF16 | L65 @ main | |
| not named on card measured by Ar3ntizcolumn alignment of row L66 uncertain [L62,L66] | MLX 8-bit (LM Studio) | MLX | — | — | — | — | — | GSM8K 17; MATH 12 | vs BF16 | L66 @ main | |
| not named on card measured by AveLabsOfficial Qwen3.8-27B NVFP4 build measured by card author [L84-91] | Stock Qwen3.8-27B NVFP4 (chat) | NVFP4 | — | — | — | — | — | GSM8K 86.0%; MMLU-Pro 75.8% | vs BF16 | L89 @ main | |
| not named on card measured by Chungulusreference perplexity on the same 106 positions [L164] | BF16 reference | — | — | — | perplexity": 9.84727010437735, | — | — | vs BF16 | L164 @ main | ||
| not named on card quoted by DuoNeural, not measuredLabeled 'Official' reference column [L72-78]; source not named | Official Uncompressed BF16 | GGUF | — | — | — | — | — | ARC-Challenge 58.87%; Winogrande 75.85%; GSM8K 90.0%; HumanEval 68.0%; HellaSwag 82.82%; TruthfulQA 77.0% | vs BF16 | L72 @ main | |
| not named on card measured by Freaksterzmeasured on same harness [L19, L28] | Minachist W8A16 AutoRound gs128 (weight-only) | AutoRound | — | 0.00230 | 97.9% | — | — | — | vs BF16 | L28 @ main | |
| not named on card measured by Freaksterzmeasured on same harness [L33] | lokeshe09 INT8 W8A8 (264 modules) | INT8 | — | 0.02110 | 94.0% | — | — | — | vs BF16 | L33 @ main | |
| not named on card quoted by JMVRoill, not measuredReported by orcarouter on separate FP8 build [L215-220] | base FP8 (orcarouter table, left of arrow) | NVFP4 | — | — | — | — | — | MMLU 84.3; MMLU-Pro 77.6; GSM8K 90.0; CMMLU 81.4 | other or no reference | L225 @ main | |
| not named on card quoted by JMVRoill, not measuredMeasured on FP8 build, not on this artifact [L215-217] | orcarouter Uncensored FP8 build (right of arrow) | NVFP4 | — | — | — | — | — | MMLU 84.7; MMLU-Pro 76.8; GSM8K 88.7; CMMLU 80.8 | other or no reference | L225 @ main | |
| not named on card measured by QTMLAMD reference release compared independently by card author [L276-284] | AMD reference MXFP4 | MXFP4 | — | 0.111 | — | — | — | — | vs BF16 | L280 @ main | |
| not named on card quoted by Reza2kn, not measuredbenchmark author's published reference, not an independent run here [L110] | Published Jev 1.13.0 reference | other | — | — | — | — | — | Jev Noul % 99.38 | other or no reference | L103 @ main | |
| not named on card measured by WarlaxZSize 17.56 GB mentioned in row label, no candidate id [L91] | Unsloth UD-Q4_K_XL (llama.cpp formats) | — | 0.0156 | — | — | — | — | other or no reference | L91 @ main | ||
| not named on card measured by abhishekchohanBase measured by card author on same harness and hardware [L58-60] | Base BF16 | — | — | — | — | — | GSM8K 95.60; GPQA Diamond 61.11; Minerva Math500 73.2; RealWorldQA 84.31; MathVision 78.06; ARC-Easy 82.53; ARC-Challenge 56.91; HellaSwag 56.50; Winogrande 76.16; TruthfulQA MC1 35.99; Average 70.04 | other or no reference | L70 @ main | ||
| not named on card measured by akopytko[L67] | unsloth NVFP4 | NVFP4 | — | 0.070824 +/- 0.000733 | 88.061 +/- 0.144% | — | — | — | vs BF16 | L67 @ main | |
| not named on card measured by akopytko[L68] | NVIDIA+Q6_K | — | 0.039314 +/- 0.000432 | 90.784 +/- 0.128% | — | — | — | vs BF16 | L68 @ main | ||
| not named on card measured by armin1Qwen's official FP8 measured by card author on GSM8K [L124, L46] | Qwen FP8 | FP8 | — | — | — | — | — | GSM8K 95.98% | vs BF16 | L124 @ main | |
| not named on card measured by avlp12measured by card author on same 35 windows: KL 0.00322, top-1 97.33% [L320]; their published 0.00216 is single-window [L321] | mlx-community 6-bit (their 6-bit) | MLX | — | 0.00322 | — | — | — | — | vs BF16 | L320 @ main | |
| not named on card measured by avlp12KL 0.00105 / top-1 98.47% on same 35 windows, a tie with this build [L324] | mlx-community 8-bit | MLX | — | 0.00105 | 98.47% | — | — | — | vs BF16 | L324 @ main | |
| not named on card measured by cdiamondsame short sample [L55] | Q4_1 | — | — | — | 6.1127 | — | — | other or no reference | L55 @ main | ||
| not named on card measured by cdiamondshort WikiText-2 control [L20] | Plain Q4_0 | — | — | — | 6.3798 | — | — | other or no reference | L20 @ main | ||
| not named on card measured by cdiamondshort WikiText-2 control [L20] | ready-made FP4 file | NVFP4 | — | — | — | 6.4949 | — | — | other or no reference | L20 @ main | |
| not named on card measured by dfed24reference is Q8_0, not bf16 [L38] | unsloth UD-Q4_K_M (llama.cpp vs unsloth Q8_0 reference) | — | — | — | — | — | kld:wiki 0.0104; top1:wiki 95.89; kld:code 0.0163; top1:code 98.75; kld:multilingual 0.0085; top1:multilingual 95.85 | vs BF16 | L43 @ main | ||
| not named on card measured by guideboardlabsrow label states 11.9GB, no table size cell [L44] | UD-IQ3_XXS (control) | — | — | — | — | — | ppl:wikitext-2 7.1553 | other or no reference | L44 @ main | ||
| not named on card quoted by intel-ai, not measuredsibling reference card numbers, cross-linked [L37,L114] | full-MXFP8 (MTP BF16) sibling | MXFP8 | — | — | — | — | — | mmlu 0.8348 | vs BF16 | L114 @ main | |
| not named on card measured by kirillbilchenkoReference row [L137] | BF16 source | — | — | — | — | — | ppl:wikitext2 7.980864 | vs BF16 | L137 @ main | ||
| not named on card measured by kirillbilchenkoThird-party MXFP4 comparison measured on same protocol [L140] | AMD Quark-AWQ MXFP4 | MXFP4 | — | — | — | — | — | ppl:wikitext2 8.187544 | vs BF16 | L140 @ main | |
| not named on card measured by lackonendesQwen3.8-27B-DFlash2-EXL3-5.0bpw comparison artifact from Mia's ExLlamaV3 fork [L221-223]; 14.2 GB in row label; EXL3 speed from different harness [L103] | MIA / EXL3 | EXL3 | — | — | — | — | — | MMLU-Pro 65.2%; IFBench_loose 34.4; IFBench_strict 32.8; HumanEval 95.7; HumanEval+ 90.9; GSM8K 93.0% | other or no reference | L93 @ main | |
| not named on card measured by lackonendes10.2 GB in row label [L96] | Escha-W2 | — | — | — | — | — | MMLU-Pro 61.8%; IFBench_loose 31.2; IFBench_strict 29.7; HumanEval 95.1; HumanEval+ 90.9; GSM8K 95.0% | other or no reference | L96 @ main | ||
| not named on card measured by lackonendes7.3 GB in row label [L99]; HumanEval base is a dash, c85 is the HumanEval+ valuebpw cell is '~2' | Unsloth IQ2XXS | — | — | — | — | — | MMLU-Pro 46.2%; IFBench_loose 23.4; IFBench_strict 14.1; HumanEval+ —<br>84.1%; GSM8K 93.0% | other or no reference | L99 @ main | ||
| not named on card measured by lackonendes9.0 GB in row label [L100]; HumanEval base is a dash, c90 is HumanEval+bpw cell is '~2' | AtomicChat AD-IQ2XXS | — | — | — | — | — | MMLU-Pro 53.0%; IFBench_loose 20.3; IFBench_strict 20.3; HumanEval+ —<br>73.8%; GSM8K 94.0% | other or no reference | L100 @ main | ||
| not named on card measured by nicosuterbaseline measured by same author on same items [L102-104] | FP8 checkpoint (baseline arm) | FP8 | — | — | — | — | — | BFCL 81.27; GPQA_Diamond 89.90; LiveCodeBench_v6 87.71; MMMU-Pro 76.91; Multimodal 86.08; macro_5_suites 84.37; RULER 92.17 | other or no reference | L112 @ main | |
| not named on card measured by oracomputingretention 92.1% (c21), body BPW c20 [L48] | Escha W2 | — | — | — | — | — | MMLU-Pro 71.24; AIME-25 87.50; Average 79.37 | other or no reference | L48 @ main | ||
| not named on card quoted by rdtand, not measureduser-reported tool-eval-bench v2.5.1 score under vLLM 0.27.2rc1 [L212-214] | PrismaQuant AQUA-AURA 5.5 bpp (community ToolEvalBench) | other | — | — | — | — | — | tool-eval-bench 93/100 | vs BF16 | L213 @ main | |
| not named on card measured by tooltdEvalPlus only [L47] | UD3-Q3_K_XL | — | — | — | — | — | HumanEval 0.823; HumanEval+ 0.805; MBPP 0.881; MBPP+ 0.751 | vs BF16 | L47 @ main | ||
| notnaurw/oQ5e-SAGE-ARCS-mtp agentic head-KL of SAGE not given as a single number in digest; stock head-KL in prose [L158]publisher comparison build | SAGE | other | — | — | — | — | — | kld:stock 0.010348; kld:agentic_fullvocab 0.434296 | vs BF16 | L158 @ main | |
| notnaurw/oQ5e-SAGE-CREST-mtp Comparator built by card author [L100,L102]publisher comparison build | uniform 6-bit (MLX affine, group 64, text-only, bf16 MTP) | other | 22.71 | download | — | — | — | — | kld:generalist_head 0.008508 | vs BF16 | L100 @ main |
| notnaurw/oQ5e-SAGE-CREST-mtp 1.54 GB smaller than SAGE-CREST [L100]publisher comparison build | SAGE-PATHS | other | — | — | — | — | — | kld:generalist_head 0.012280 | vs BF16 | L197 @ main | |
| notnaurw/oQ5e-SAGE-CREST-mtp generalist head-KL 0.010153 [L197]publisher comparison build | SAGE | other | — | — | — | — | — | kld:generalist_head 0.010153 | vs BF16 | L197 @ main | |
| notnaurw/oQ5e-SAGE-PATHS-mtp comparator built by the card author [L140]publisher comparison build | uniform 6-bit (MLX affine, group 64, text-only) | other | 22.71 | download | — | — | — | — | kld:generalist_head 0.008508 | vs BF16 | L140 @ main |
| notnaurw/oQ5e-SAGE-PATHS-mtp generalist head-KL 0.010153 prose only [L645]; 5.92 bpw prose only [L104]publisher comparison build | SAGE | other | — | — | — | — | — | kld:generalist_head 0.010153 | vs BF16 | L645 @ main | |
| notnaurw/oQ5e-SAGE-PATHS-mtp generalist head-KL 0.014043 prose only [L197]publisher comparison build | SAGE-ARCS | other | — | — | — | — | — | kld:generalist_head 0.014043 | vs BF16 | L197 @ main | |
| nova-quill/OBLITERATUS-OBLITERATED-unsloth-mlx earlier version, n=285 [L115]publisher comparison build | V1 (s51, aggressive) | MLX | — | — | — | — | — | MMLU 81.4% | other or no reference | L115 @ main | |
| nvidia/NVFP4 measured by juzbox[L35,L69,L84] | nvidia/Qwen3.8-27B-NVFP4 | NVFP4 | 21.92 | download | 0.0362 | — | — | — | BFCL v3 all (custom engine) 90.1%; BFCL parallel (custom engine) 79.0% | vs BF16 | L35 @ main |
| nvidia/NVFP4 measured by gittensor-model-hubMeasured by this card's author on same box [L156-158]; AA-LCR value for NVIDIA not in candidates (text says all three 55 [L153 alt]); 95% CI overall [73.9-82.7] c18 | NVIDIA NVFP4 | NVFP4 | 21.00 | download | — | — | — | — | GPQA Diamond 76.8%; MMLU-Pro 83.0%; Overall (328) 78.7% | reference not stated | L162 @ main |
| nvidia/NVFP4 measured by Triodymeasured by card author [L109] | nvidia/Qwen3.8-27B-NVFP4, SGLang's recipe | NVFP4 | — | — | — | — | — | GSM8K_lenient 97.25%; GSM8K_strict 95.25%; HumanEval_lenient 90.85%; HumanEval_strict 86.59% | vs BF16 | L109 @ main | |
| nvidia/NVFP4 measured by armin1[L90] | nvidia/Qwen3.8-27B-NVFP4 | NVFP4 | — | 0.0244 | 94.89% | — | — | kld_99 0.272 | vs BF16 | L90 @ main | |
| nvidia/NVFP4 measured by armin1Measured by card author on same documents [L112] | nvidia/Qwen3.8-27B-NVFP4 | NVFP4 | — | 0.0244 | 94.89% | — | — | kld_99 0.272 | vs BF16 | L112 @ main | |
| NVIDIA NVFP4 release (repo link truncated on card digest; nvidia/Qwe...) measured by atakansMeasured by card author with same script [L25] | NVFP4 (NVIDIA) | NVFP4 | 22.55 | download | 0.0398 | 88.7% | — | 1.029 | kld_99 0.348 | vs BF16 | L31 @ main |
| Omega-Factory/GGUF size approximate (~) [L33]publisher comparison build | FP16 Baseline | GGUF | 58.00 | download | — | — | — | — | ARC-MC 94.5% (189/200); GSM8K 59.0% (118/200); MATH-500 23.0% (46/200) | other or no reference | L33 @ main |
| oracomputing/OQ3-GGUF reference, retention 100% [L46]publisher comparison build | Qwen3.8-27B (bf16) | GGUF | — | — | — | — | — | MMLU-Pro 76.71; AIME-25 95.83; Average 86.27 | other or no reference | L46 @ main | |
| orcarouter/OrcaSAQ-2-27B reference row [L118]publisher comparison build | Qwen3.8-27B BF16 | other | 54.00 | download | — | 100% | 5.6468 | — | — | vs BF16 | L118 @ main |
| OS-Software/Uncensored-Heretic-v2-GGUF KL 0 by definition [L33]publisher comparison build | Original Qwen/Qwen3.8-27B (reference) | GGUF | — | 0 *(by definition)* | — | — | — | — | other or no reference | L33 @ main | |
| penkia/TernaryQuench-GGUF reference row measured with MLX [L60-61]publisher comparison build | Qwen3.8-27B 4-bit | GGUF | — | — | — | — | — | ppl:wikitext2 7.7471; ppl_second_half:wikitext2 6.2370 | vs BF16 | L54 @ main | |
| penkia/TernaryQuench-GGUF reference rows use native PyTorch [L79]publisher comparison build | Qwen3.8-27B BF16 | GGUF | — | — | — | — | — | PIQA 81.61; ARC-Easy 72.98; ARC-Challenge 58.87; HellaSwag 82.93; WinoGrande 75.37 | vs BF16 | L71 @ main | |
| perplexity-ai/pplx-decider-v1-27b measured by AveLabsOriginal BF16 decider measured by card author on same items [L62-68] | pplx-decider BF16 (decisions) | — | — | — | — | — | WANLI 68.1% | vs BF16 | L66 @ main | ||
| poisonxa/PXQN-OneCard [L107]publisher comparison build | Classic PXQ4 (our earlier format) | other | 16.52 | download | 0.0176 | 96.0% | — | — | ppl:wikitext-2 6.816 | vs a Q8_0 | L107 @ main |
| poisonxa/PXQN-OneCard [L105]publisher comparison build | Classic PXQ3 (our earlier format, same size class) | other | 12.66 | download | 0.2033 | 88.5% | — | — | ppl:wikitext-2 7.371 | vs a Q8_0 | L105 @ main |
| PollardWeights/Penjing-27B-Pollard f16 reference PPL 4.1422 [L64]; size 54.64 GB from prose headline [L26]publisher comparison build | f16 reference | other | 54.64 | download | — | — | — | — | ppl:wikitext2 PPL 4.1422._ | other or no reference | L26 @ main |
| prism-ml/Ternary-Bonsai-2-27B-gguf measured by oshitekuMeasured by card author with same prompts and settings [L66] | Ternary Bonsai 2 | GGUF | 7.2 | download | 0.153 | 89.1% | — | — | SWE-bench Verified Mini 60% | vs BF16 | L70 @ main |
| prism-ml/Ternary-Bonsai-2-27B-gguf measured by dealignaiBase ternary quant measured by card author for comparison [L106] | Base PQ2_0 | GGUF | — | — | — | — | — | MMLU 40.53% | other or no reference | L106 @ main | |
| pugant/MTP-Q3_0_ROCMFPX control reproduced within sigma [L76]; PPL and delta share one cell [L80]publisher comparison build | STRIX_LEAN Q4_0_ROCMFP4_STRIX_LEAN (baseline) | GGUF | 14.84 | download | — | — | — | — | ppl:en 6.8226 (+2.74%); ppl:it 12.1168 (+3.42%); ppl_delta_pct:en 6.8226 (+2.74%); ppl_delta_pct:it 12.1168 (+3.42%) | vs BF16 | L80 @ main |
| pugant/MTP-Q3_0_ROCMFPX historical reference values [L85]publisher comparison build | BF16 baseline | GGUF | — | — | — | — | — | ppl:en 6.6409; ppl:it 11.7156 | vs BF16 | L75 @ main | |
| QTML/Security-Quark-MXFP4 BF16 -> MXFP4 KL 0.116; top-1 token flips 10% not mapped (flips, not agreement) [L119, L280]KL instrument differs from the 0.0151 figure [L128]publisher comparison build | Base Qwen3.8-27B MXFP4 | MXFP4 | — | 0.116 | — | — | — | — | vs BF16 | L280 @ main | |
| quark75/Q4_0.gguf reference perplexity [L26]publisher comparison build | BF16 original | GGUF | — | — | — | 6.9508 +/- 0.0449 | — | — | vs BF16 | L26 @ main | |
| QUASAR-QAT/QUASAR-NVFP4 measured by juzboxengine uses NVIDIA NVFP4 lm_head for QUASAR in custom engine runs [L79-80] | QUASAR-QAT/Qwen3.8-27B-QUASAR-NVFP4 | NVFP4 | 20.56 | download | 0.0475 | — | — | — | BFCL v3 all (custom engine) 87.0%; BFCL parallel (custom engine) 65.5% | vs BF16 | L35 @ main |
| QUASAR-QAT/QUASAR-NVFP4 measured by gittensor-model-hubMeasured by this card's author on same box [L156-158]; AIME 2025 cell not in candidates; 95% CI overall [75.5-84.1] c17 | QUASAR NVFP4 | NVFP4 | 19.70 | download | — | — | — | — | GPQA Diamond 78.8%; MMLU-Pro 84.0%; Overall (328) 80.2%; AA-LCR 55 | reference not stated | L162 @ main |
| QUASAR-QAT/QUASAR-NVFP4 [L91]publisher comparison build | GPTQ NVFP4 (LLM Compressor, ours) | NVFP4 | — | — | — | — | — | MRCR 61.3; MRCR 8-needle 35.5; HMMT'26 69.7; SuperGPQA 58.9; LiveCodeBench 77.9; RULER 32K 96.0; RULER 64K 95.4; RULER 128K 95.3 | vs BF16 | L91 @ main | |
| QUASAR-QAT/QUASAR-NVFP4 [L92]publisher comparison build | RTN NVFP4 (LLM Compressor, ours) | NVFP4 | — | — | — | — | — | MRCR 57.3; MRCR 8-needle 32.8; HMMT'26 66.9; SuperGPQA 58.6; LiveCodeBench 77.9; RULER 32K 96.2; RULER 64K 96.1; RULER 128K 95.3 | vs BF16 | L92 @ main | |
| Qwen/Qwen3.8-27B measured by cloudnathan5BF16 base measured on same settings [L93,L98] | Qwen/Qwen3.8-27B (BF16) | 55.60 | download | — | — | — | — | ppl:wikitext2 6.5574 | vs BF16 | L29 @ main | |
| Qwen/Qwen3.8-27B measured by ThakiCloudReference measured by card author on same harness [L106, L141] | Qwen/Qwen3.8-27B (bf16) | 55.59 | download | — | — | — | — | MMMU-val MC 0.8455 | vs BF16 | L36 @ main | |
| Qwen/Qwen3.8-27B measured by ThakiCloudbf16 reference measured by card author [L106,L136]Earlier bf16 MMMU 0.4511 withdrawn as a broken measurement [L115] | Qwen/Qwen3.8-27B (bf16) | 55.59 | download | — | — | — | — | MMMU-val MC 0.8455; HAERAE 0.7709; KoBEST 0.6770; belebele-ko 0.9000; ARC-Challenge 0.5700; IFEval prompt-strict 0.3050 | vs BF16 | L36 @ main | |
| Qwen/Qwen3.8-27B measured by primitive-aireference row [L81] | Qwen BF16 | 55.51 | download | — | — | — | — | overall 89.3; knowledge 91.0; call 85.6; abstain 55.0 | vs BF16 | L81 @ main | |
| Qwen/Qwen3.8-27B measured by kacaforyah16-bit reference [L34] | Qwen3.8-27B 16-bit original | 54.70 | download | — | — | — | 1.000x | — | other or no reference | L34 @ main | |
| Qwen/Qwen3.8-27B measured by Ar3ntizPPL from 'original -> SUC1' column [L49]; repo not named on card | original Qwen3.8-27B (BF16) | 53.79 | download | — | — | 7.05 | — | — | vs BF16 | L26 @ main | |
| Qwen/Qwen3.8-27B measured by Capicua25xbf16 served by a cloud provider [L168] | ref (bf16, cloud provider) | — | — | — | — | — | GSM8K think flex 0.96; GSM8K think strict 0.82; GSM8K nothink flex 0.98; GSM8K nothink strict 0.98; IFEval inst 0.97; IFEval prompt 0.95; GPQA-Diamond 0.78; AIME'25 0.93; tau2-telecom 0.939; tau2-airline 0.760; tau2-retail 0.850; MMLU-Pro 0.804; Terminal-Bench hard 0.273 | vs BF16 | L173 @ main | ||
| Qwen/Qwen3.8-27B measured by ISTA-DASLabmeasured by card authors with same xhigh setting [L38] | Base Model | — | — | — | — | — | AIME 2025 100.00; GPQA Diamond 89.90 | other or no reference | L44 @ main | ||
| Qwen/Qwen3.8-27B measured by MiawTeamauthor-measured reference in same session [L152-153] | Qwen3.8-27B base (bf16) | — | — | — | — | — | MMLU 83.4; ARC-Challenge 58.9; HellaSwag 82.8; Winogrande 76.1; kld:first-token-vs-base 0 | other or no reference | L146 @ main | ||
| Qwen/Qwen3.8-27B measured by PilcothinkBase measured by card author [L60-63] | Qwen3.8-27B | — | — | — | — | — | MMLU 83.49%; GSM8K flexible 72.86%; GSM8K strict 70.36% | other or no reference | L62 @ main | ||
| Qwen/Qwen3.8-27B measured by ProCreationsTeacher BF16 measured on same table [L42,L73] | Original Qwen3.8 BF16 | — | — | — | — | — | ppl:wikitext2 6.5137; ARC-C 59.30; ARC-E 74.92; HellaSwag 82.15; Winogrande 75.14 | vs BF16 | L42 @ main | ||
| Qwen/Qwen3.8-27B measured by Triodybf16 reference measured by card author [L111] | Qwen/Qwen3.8-27B bf16, no speculation | — | — | — | — | — | GSM8K_lenient 97.75%; GSM8K_strict 96.75%; HumanEval_lenient 94.51%; HumanEval_strict 92.07% | vs BF16 | L111 @ main | ||
| Qwen/Qwen3.8-27B measured by Zynerjiauthor-measured baseline [L111] | pristine Qwen3.8-27B (before) | — | — | — | — | — | MMLU-val 0.812 | other or no reference | L111 @ main | ||
| Qwen/Qwen3.8-27B quoted by acyildirimer, not measuredreference row [L80] | Qwen/Qwen3.8-27B (BF16) | NVFP4 | — | — | — | — | — | MMLU 85.3; ARC-C 96.8; HellaSwag 94.3; GSM8K 97.4; HumanEval 93.9; 4-task mean 93.5 | vs BF16 | L80 @ main | |
| Qwen/Qwen3.8-27B measured by amdauthor-measured reference [L36] | BF16 base | — | — | — | 8.4364 ppl:wikitext | — | GSM8K-thinking-flexible 93.33; GSM8K-thinking-strict 93.33; GSM8K-nothink-flexible 90.67; GSM8K-nothink-strict 89.76 | vs BF16 | L29 @ main | ||
| Qwen/Qwen3.8-27B measured by cloudnathan5Reference row, same settings [L94, L99] | Qwen/Qwen3.8-27B (BF16) | — | — | — | — | — | ppl:wikitext-2 6.5574 | vs BF16 | L99 @ main | ||
| Qwen/Qwen3.8-27B measured by huginnforkBaseline row [L68] | Qwen/Qwen3.8-27B (bf16 baseline) | — | — | — | 6.9416 | — | kld:weight_only 0 | vs BF16 | L68 @ main | ||
| Qwen/Qwen3.8-27B quoted by intel-ai, not measuredupstream-reported BF16 table, not re-run on this stack [L99, L276] | BF16 Qwen/Qwen3.8-27B | MXFP8 | — | — | — | — | — | gsm8k_strict 0.9742; mmlu 0.8340; piqa_acc 0.8096; hellaswag_acc 0.6369; AVG 0.8137 | vs BF16 | L105 @ main | |
| Qwen/Qwen3.8-27B quoted by intel-ai, not measuredupstream-reported table, not a re-run on this stack [L247,L113] | BF16 baseline | MXFP8 | — | — | — | — | — | mmlu 0.8340 | vs BF16 | L113 @ main | |
| Qwen/Qwen3.8-27B measured by kirillbilchenkoReference row, measured by card author [L140, L179] | BF16 source | — | — | — | — | — | ppl:wikitext2_prompt 7.980864 | vs BF16 | L140 @ main | ||
| Qwen/Qwen3.8-27B quoted by lackonendes, not measuredPublisher numbers, not run by card author [L57-58] | BF16 Base (published) | GGUF | — | — | — | — | — | AIME 2026 98.13; GPQA-Diamond 89.93; MMLU-Pro 85.54; LiveCodeBench v6 91.14; IFBench 79.75 | vs BF16 | L62 @ main | |
| Qwen/Qwen3.8-27B measured by malaiwahBF16 reference for the MMLU-Pro pairing [L741] | Qwen/Qwen3.8-27B BF16 | — | — | — | — | — | MMLU-Pro 57/70 (81.4 %) | vs BF16 | L741 @ main | ||
| Qwen/Qwen3.8-27B quoted by neroued, not measuredcopied from upstream card; sampling settings and IFBench level not stated, not same-protocol [L321-323] | Official Qwen3.8-27B BF16 | NVFP4 | — | — | — | — | — | GPQA-Diamond 89.2; ERQA 65.5; RealWorldQA 85.9 | vs BF16 | L310 @ main | |
| Qwen/Qwen3.8-27B measured by nova-quillmeasured by card author, n=2,850, stderr +/-0.65 [L114] | Stock Qwen3.8-27B | — | — | — | — | — | MMLU 84.60% | other or no reference | L114 @ main | ||
| Qwen/Qwen3.8-27B quoted by philbert440, not measuredofficial base-model BF16 benchmarks copied from Qwen card; not re-measured on this checkpoint [L90-92]Qwen3.6-27B comparison column omitted | Qwen3.8-27B (BF16 base) | AWQ | — | — | — | — | — | Terminal Bench 2.1 73.0; SWE-bench Pro 53.5; GPQA Diamond 89.2; HLE 30.8; LiveCodeBench v6 83.9; IFBench 69.1; OmniDocBench 1.5 91.1; RealWorldQA 85.9; ERQA 65.5; OSWorld-Verified 63.9 | reference not stated | L96 @ main | |
| Qwen/Qwen3.8-27B measured by quark75[L30] | BF16 original | — | — | — | 7.002 | — | — | vs BF16 | L30 @ main | ||
| Qwen/Qwen3.8-27B quoted by stosc, not measuredTaken from upstream card; sampling settings and IFBench metric level not declared there, so not a same-protocol comparison [L273] | Official Qwen3.8-27B BF16 | NVFP4 | — | — | — | — | — | IFBench 79.5; GPQA-Diamond 89.2; ERQA 65.5; RealWorldQA 85.9 | vs BF16 | L264 @ main | |
| Qwen/Qwen3.8-27B quoted by sunnyyy, not measuredfrom upstream model card; sampling and IFBench level not stated, not same-protocol [L198-200]upstream reports no AIME [L202] | Official Qwen3.8-27B BF16 | other | — | — | — | — | — | GPQA-Diamond 89.2; ERQA 65.5; RealWorldQA 85.9 | vs BF16 | L187 @ main | |
| Qwen/Qwen3.8-27B quoted by sunnyyy, not measuredcopied from upstream card; not a same-protocol comparison; upstream reports no AIME [L284-288] | Official Qwen3.8-27B BF16 | NVFP4 | — | — | — | — | — | GPQA-Diamond 89.2; ERQA 65.5; RealWorldQA 85.9 | other or no reference | L273 @ main | |
| Qwen/Qwen3.8-27B quoted by vrfai, not measuredPublished Qwen figures for bf16 weights [L75-76]; base-card benchmark tables below L155 were not measured on this quant | Qwen card | FP8 | — | — | — | — | — | MathVision 90.0; ERQA 65.5 | vs BF16 | L75 @ main | |
| Qwen/Qwen3.8-27B quoted by vrfai, not measuredfigures copied from Qwen card [L75-76]Tables from L351 onward are base-model bf16 results reproduced verbatim, not measured on this checkpoint [L155-159]; omitted | Qwen card (published bf16 figures) | FP8 | — | — | — | — | — | MathVision 90.0; ERQA 65.5 | vs BF16 | L75 @ main | |
| Qwen/Qwen3.8-27B quoted by vrfai, not measuredQwen published figures shown beside own measurements [L79-85]; the base-card VL table (L357-370) is reproduced verbatim from Qwen and not measured on this quant [L161-165] | Qwen3.8-27B (Qwen card published) | NVFP4 | — | — | — | — | — | MathVision 90.0; ERQA 65.5 | vs BF16 | L81 @ main | |
| Qwen/Qwen3.8-27B quoted by wiklif, not measuredQwen-published figures; token budget and LCB window not stated by Qwen [L99-100] | Qwen card (BF16) | GGUF | — | — | — | — | — | GPQA Diamond 89.2; IFBench 79.5; LiveCodeBench v6 90.3 | vs a Q8_0 | L67 @ main | |
| Qwen/FP8 measured by TelperionAIofficial FP8 release as reference; card excludes it from comparison [L60] | Qwen/Qwen3.8-27B-FP8 (8-bit ref) | FP8 | 30.90 | download | — | 96.15% | — | — | — | vs BF16 | L54 @ main |
| Qwen/FP8 measured by TelperionAI8-bit reference [L58] | Qwen/Qwen3.8-27B-FP8 | FP8 | 30.90 | download | — | 96.15% | — | — | near-tie disagree 22.70%; moderate disagree 3.48%; confident disagree 1.45%; certain disagree 0.08%; divmed 47 | vs BF16 | L58 @ main |
| Qwen/FP8 measured by TelperionAIreference at different precision [L56,L62] | Qwen/Qwen3.8-27B-FP8 (8-bit ref) | FP8 | 30.90 | download | — | 96.15% | — | — | divmed 47 | vs BF16 | L56 @ main |
| Qwen/FP8 measured by TelperionAI[L56] | Qwen/Qwen3.8-27B-FP8 (8-bit ref) | FP8 | 30.90 | download | — | 96.15% | — | — | — | vs BF16 | L56 @ main |
| Qwen/FP8 measured by TelperionAIreference at different precision, excluded from comparison [L62] | Qwen/Qwen3.8-27B-FP8 (8-bit ref) | FP8 | 30.90 | download | — | 96.15% | — | — | — | vs BF16 | L56 @ main |
| Qwen/FP8 measured by luedmeasured by card author; FP8 weights dequantized to BF16 in HF path [L187]size is weight files [L90] | Official Qwen FP8 / BF16 A16 | FP8 | 30.87 | download | 0.004396 | 98.53% | — | — | — | vs BF16 | L90 @ main |
| Qwen/FP8 measured by shisa-aiMeasured locally by card author at revision 017b9c7a; median KLD c11 [L56] | Qwen3.8-27B-FP8 (official) | FP8 | 30.87 | download | 0.005189 | 96.93% | — | — | kld_99 0.052687 | vs BF16 | L56 @ main |
| Qwen/FP8 measured by malaiwahshard-0 [L460] | official FP8 (shard-0 replay) | FP8 | 30.61 | resident | 0.005197 | 96.92 % | — | — | — | vs BF16 | L460 @ main |
| Qwen/FP8 measured by malaiwahMMLU-Pro 56/70 [L745]v3 value c62 0.013126 on v3 suite [L395] | Qwen/Qwen3.8-27B-FP8 | FP8 | 30.61 | resident | 0.005294 | 96.79 % | — | — | kld:v4 0.005891; top1:v4 96.72 %; MMLU-Pro 56/70 (80.0 %); kld:shard0 0.005197; top1:shard0 96.92 %; kld:v3 0.013126 | vs BF16 | L237 @ main |
| Qwen/FP8 measured by malaiwah[L323] | Qwen/Qwen3.8-27B-FP8 v5 10M | FP8 | 30.61 | resident | 0.005294 | 96.79 % | — | — | — | other or no reference | L323 @ main |
| Qwen/FP8 measured by malaiwah[L500] | official FP8 vLLM shard0 | FP8 | 30.61 | resident | 0.005197 | 96.92 % | — | — | — | other or no reference | L500 @ main |
| Qwen/FP8 measured by malaiwah[L781] | Qwen/Qwen3.8-27B-FP8 v3 136-context | FP8 | 30.61 | resident | 0.013126 | 96.22 % | — | — | — | other or no reference | L781 @ main |
| Qwen/FP8 measured by Capicua25xAA-LCR cell holds two seeds (s1234 / s99) [L176]FP8 repo named at L73 | B FP8 @ fp8 KV | FP8 | — | — | — | — | — | GSM8K think flex 0.98; GSM8K think strict 0.86; IFEval inst 0.98; IFEval prompt 0.98; GPQA-Diamond 0.85; AIME'25 0.97; tau2-telecom 0.94; tau2-airline 0.86; tau2-retail 0.82; MMLU-Pro 0.817; Terminal-Bench hard 0.341 | vs BF16 | L173 @ main | |
| Qwen/FP8 measured by Freaksterzrepo named in prose [L10]; measured on same harness [L30] | Qwen official FP8 (block-128) | FP8 | — | 0.00584 | 96.9% | — | — | — | vs BF16 | L30 @ main | |
| Qwen/FP8 measured by armin1[L88,L100] | Qwen/Qwen3.8-27B-FP8 | FP8 | — | 0.0034 | 98.22% | — | — | kld_99 0.037; GSM8K 95.98% | vs BF16 | L88 @ main | |
| Qwen/FP8 measured by dancinlabMeasured by card author with same scripts [L222] | Base FP8 | FP8 | — | — | — | — | — | MMLU 84.3%; MMLU-Pro 77.6%; GSM8K 90.0%; CMMLU 81.4% | other or no reference | L225 @ main | |
| Qwen/FP8 measured by gubernac[L65,L159] | Base FP8 | FP8 | — | — | — | — | — | MMLU 84.3%; MMLU-Pro 77.6%; GSM8K 90.0%; CMMLU 81.4% | other or no reference | L208 @ main | |
| Qwen/FP8 measured by huginnforkHand-dequantised to evaluate on same harness [L125-128]407 modules quantised [L69] | Qwen/Qwen3.8-27B-FP8 (official) | FP8 | — | — | — | 6.9539 | +0.18 % | kld:weight_only 0.0523; kld:as_deployed 0.1001 | vs BF16 | L69 @ main | |
| Qwen/FP8 measured by huginnforkmean_kld is weight-only column, kld_as_deployed is the other column; official checkpoint hand-dequantised for evaluation [L125-128]407 modules quantised [L69] | Qwen/Qwen3.8-27B-FP8 | FP8 | — | 0.0523 | — | 6.9539 | +0.18 % | kld_as_deployed 0.1001 | vs BF16 | L69 @ main | |
| Qwen/FP8 measured by malaiwah[L856] | Qwen/Qwen3.8-27B-FP8 v4 36-context | FP8 | — | 0.005891 | 96.72 % | — | — | — | other or no reference | L856 @ main | |
| Qwen (link truncated in prompt) measured by primitive-ai[L82] | Qwen FP8 | FP8 | 30.82 | download | — | — | — | — | overall 89.6; knowledge 91.4; call 85.0; abstain 55.0 | vs BF16 | L82 @ main |
| Qwen FP8 release (repo link truncated on card digest) measured by atakansMeasured by card author with same script [L25] | FP8 (Qwen) | FP8 | 31.14 | download | 0.0046 | 96.5% | — | 1.002 | kld_99 0.045 | vs BF16 | L29 @ main |
| RadixArk/NVFP4 measured by shisa-aiMeasured locally at revision 52d1adc5; median KLD c26; two captures agreed within 3.3e-5 mean KLD [L59, L62] | RadixArk/Qwen3.8-27B-NVFP4 | NVFP4 | 21.93 | download | 0.041985 | 91.55% | — | — | kld_99 0.465240 | vs BF16 | L59 @ main |
| RadixArk/NVFP4 measured by TelperionAI[L58] | RadixArk/Qwen3.8-27B-NVFP4 | NVFP4 | 21.90 | download | — | 90.23% | — | — | divmed 11 | vs BF16 | L58 @ main |
| RadixArk/NVFP4 measured by TelperionAI | RadixArk/Qwen3.8-27B-NVFP4 | NVFP4 | 21.90 | download | — | 90.23% | — | — | — | vs BF16 | L58 @ main |
| RadixArk/NVFP4 measured by armin1[L92] | RadixArk/Qwen3.8-27B-NVFP4 | NVFP4 | — | 0.0301 | 94.41% | — | — | kld_99 0.330 | vs BF16 | L92 @ main | |
| RadixArk/NVFP4-BF16-LMHead measured by TelperionAIRTN, identical precision map [L59,L93-95] | RadixArk/Qwen3.8-27B-NVFP4-BF16-LMHead | NVFP4 | 23.80 | download | — | 91.06% | — | — | — | vs BF16 | L59 @ main |
| RadixArk/NVFP4-BF16-LMHead measured by armin1[L91,L98] | RadixArk/Qwen3.8-27B-NVFP4-BF16-LMHead | NVFP4 | — | 0.0245 | 95.36% | — | — | kld_99 0.302; GSM8K 96.21% | vs BF16 | L91 @ main | |
| RadixArk/NVFP4-BF16-LMHead measured by armin1Measured by card author [L113, L122] | RadixArk/Qwen3.8-27B-NVFP4-BF16-LMHead | NVFP4 | — | 0.0245 | 95.36% | — | — | kld_99 0.302; GSM8K 96.21% | vs BF16 | L113 @ main | |
| rapid-mlx/mixed-3.5bpw-MLX identical-subset comparison, 95% CI about +-7 points at n=200 [L118-121]publisher comparison build | bf16 (subset) | MLX | 51.00 | download | — | — | — | — | MMLU-Pro (n=200) 59.5; GSM8K (n=60) 90.0; MMBench-EN (n=150) 88.0 | other or no reference | L125 @ main |
| rapid-mlx/mixed-3.5bpw-MLX identical-subset comparison [L123-127]publisher comparison build | 4-bit (subset) | MLX | 16.00 | download | — | — | — | — | MMLU-Pro (n=200) 61.5; GSM8K (n=60) 98.3; MMBench-EN (n=150) 92.0 | other or no reference | L125 @ main |
| rdtand/PrismaAQUA-gridbook-13GB-5080-vllm measured through same tool, corpus, seqlen [L356-359]publisher comparison build | BF16 reference | other | — | — | — | — | — | ppl:wikitext2 9.365 | vs BF16 | L328 @ main | |
| rdtand/PrismaScout-AQUA-Vision-20GB reference PPL [L119]publisher comparison build | BF16 source | other | — | — | — | 9.366213045 | — | — | vs BF16 | L119 @ main | |
| rdtand/PrismaScout-AQUA-Vision-20GB sibling rdtand/Qwen3.8-27B-PrismaScout-AQUA-20GB [L154,L161,L431]Qwen3.6 PrismaScout figures (0.0779318074 KL) are a different model, omitted [L138-139]publisher comparison build | PrismaScout-AQUA-20GB (visionless sibling) | other | — | 0.0402201669 | — | — | — | — | vs BF16 | L154 @ main | |
| rdtand (link truncated in prompt) measured by primitive-ai[L83] | rdtand PrismaAQUA 5.5bit | 23.62 | download | — | — | — | — | overall 89.4; knowledge 91.2; call 85.6; abstain 52.5 | vs BF16 | L83 @ main | |
| RedHatAI/NVFP4 quoted by mlx-bun, not measuredBaselines from Red Hat's published evaluation, linked at the RedHatAI/Qwen3.8-27B-NVFP4 evaluation section [L111]; approx BF16 size 54 GB from prose [L67] | BF16 base | MLX | 54.00 | download | — | — | — | — | GSM8K-Platinum_strict 96.25%; GPQA Diamond 89.56% | vs BF16 | L67 @ main |
| RukaRat/INT8-W8A8-imatrix-heretic-MTP same battery, same box [L43-47]refusal rates harmful 96% off c22, 94% on c23 [L68]publisher comparison build | INT8 W8A8 (unmodified) | INT8 | 31.25 | download | — | — | — | — | quality/160 132 | other or no reference | L47 @ main |
| RukaRat/INT8-W8A8-imatrix-heretic-MTP [L51]publisher comparison build | W4A16 plain, abliterated | INT8 | 19.43 | download | — | — | — | — | quality/160 135 | other or no reference | L51 @ main |
| RukaRat/INT8-W8A8-imatrix-heretic-MTP [L49]publisher comparison build | W4A16 + int8 head | INT8 | 17.82 | download | — | — | — | — | quality/160 131 | other or no reference | L49 @ main |
| RukaRat/INT8-W8A8-imatrix-heretic-MTP [L50]publisher comparison build | W4A16 + int8 head, abliterated | INT8 | 17.82 | download | — | — | — | — | quality/160 134 | other or no reference | L50 @ main |
| RukaRat/W4A16-imatrix-heretic-int8head-MTP Size 18.1 GiB c21 [L179]; L60 table shows 131/160 c1.0 for this build in an earlier comparison, conflicts with 135 here, not reconciledpublisher comparison build | W4A16 plain, abliterated | INT8 | — | — | — | — | — | five_task_battery_correct_of_160 135 | other or no reference | L219 @ main | |
| RukaRat/W4A16-imatrix-heretic-int8head-MTP Size 16.6 GiB c15 [L177]publisher comparison build | W4A16 + int8 head (non-abliterated) | INT8 | — | — | — | — | — | five_task_battery_correct_of_160 131 | other or no reference | L217 @ main | |
| RukaRat/W4A16-imatrix-heretic-int8head-MTP MTP accept think not available (dash) [L220]publisher comparison build | W4A16 plain | INT8 | — | — | — | — | — | five_task_battery_correct_of_160 131 | other or no reference | L220 @ main | |
| RukaRat/W4A16-imatrix-heretic-int8head-MTP Size 29.1 GiB c10 [L175]publisher comparison build | INT8 W8A8 | INT8 | — | — | — | — | — | five_task_battery_correct_of_160 132 | other or no reference | L215 @ main | |
| RukaRat/W4A16-imatrix-heretic-int8head-MTP Size 29.1 GiB c13 [L176]publisher comparison build | INT8 W8A8, abliterated | INT8 | — | — | — | — | — | five_task_battery_correct_of_160 130 | other or no reference | L216 @ main | |
| sakamakismile/MTP-NVFP4 measured by malaiwahshard-0 [L465] | sakamakismile/Qwen3.8-27B-MTP-NVFP4 | NVFP4 | 20.56 | download | 0.151280 | 84.74 % | — | — | — | vs BF16 | L465 @ main |
| sakamakismile/MTP-NVFP4 measured by malaiwahshard-0 [L474] | sakamakismile/Qwen3.8-27B-MTP-NVFP4 | NVFP4 | 20.56 | download | 0.151280 | 84.74 % | — | — | — | vs BF16 | L474 @ main |
| sakamakismile/MTP-NVFP4 measured by malaiwah[L505] | sakamakismile/Qwen3.8-27B-MTP-NVFP4 shard0 | NVFP4 | 20.56 | download | 0.151280 | 84.74 % | — | — | — | other or no reference | L505 @ main |
| sammyblues/IQ2_XS-GGUF Layer-pruned negative experiment, not published [L124-129]publisher comparison build | 48L IQ2_XS experiment | GGUF | 7.1 | download | — | — | — | — | ppl:wikitext2 144,585.2858; HellaSwag-100 22% | vs BF16 | L109 @ main |
| sammyblues/IQ2_XS-GGUF Layer-pruned negative experiment, not published [L124-129]publisher comparison build | 40L IQ2_XS experiment | GGUF | 6.2 | download | — | — | — | — | ppl:wikitext2 8,201.4746; HellaSwag-100 26% | vs BF16 | L110 @ main |
| shisa-ai/FP8-BLOCK Publisher's own comparison build; median KLD c16 [L57]publisher comparison build | Qwen3.8-27B-FP8-BLOCK-GDN8 (official-policy clone) | FP8 | 30.87 | download | 0.005253 | 96.86% | — | — | kld_99 0.053828 | vs BF16 | L57 @ main |
| SmokeyPete/Heretic-NVFP4-DGX-Spark reference is the abliterated BF16, not original Qwen [L108]first-token KL 0.061 on harmless prompts is for the abliteration edit, not the quant [L200]publisher comparison build | BF16 (abliterated, pre-quantization) | NVFP4 | — | — | — | — | — | GSM8K 97.12%; HumanEval+ 92.07% | vs BF16 | L112 @ main | |
| soyrsoyr/W4A16-AWQ-GPTQ reference scored in same run [L102]publisher comparison build | Qwen3.8-27B bf16 base | AWQ | — | — | — | — | — | OpenLLM v1 average 65.20; Reasoning average 90.11 | vs BF16 | L81 @ main | |
| staccs/lecore-GGUF numbers measured on the companion MLX export, not the GGUF [L42,L10]subsets, expect about 2-4 points sampling noise [L38]publisher comparison build | leCore MLX export | GGUF | — | — | — | — | — | MMLU-5shot 84.9%; GSM8K-strict 91.0%; GSM8K-flexible 89.0% | reference not stated | L42 @ main | |
| staccs/lecore-GGUF simulated comparison build, same subsets [L43]publisher comparison build | stock 3-bit affine, group 64 (simulated) | GGUF | — | — | — | — | — | MMLU-5shot 81.4%; GSM8K-strict 60.0%; GSM8K-flexible 62.0% | reference not stated | L43 @ main | |
| staccs/lecore-mlx Benchmark row is labelled 'simulated' [L39]publisher comparison build | stock MLX 3-bit, group 64 | MLX | 11.77 | text weights (excl. vision, MTP) | — | — | — | — | ppl:wikitext2 9.2464; ppl:python-code 3.5006; MMLU 5-shot 81.4%; GSM8K strict 60.0%; GSM8K flexible 62.0% | vs BF16 | L31 @ main |
| staccs/lecore-mlx publisher comparison build | bf16 reference | MLX | — | — | — | — | — | ppl:wikitext2 7.8880; ppl:python-code 2.5651 | vs BF16 | L29 @ main | |
| tcclaviger/RFI8-AA reference row [L37]publisher comparison build | BF16 original | other | — | — | — | 6.9224 | — | — | vs BF16 | L37 @ main | |
| TelperionAI/EXL3-4.0bpw publisher's own comparison build [L29]publisher comparison build | TelperionAI INT4-AWQ-GPTQ (vLLM) | EXL3 | 25.10 | download | — | 96.30% | — | — | — | vs BF16 | L29 @ main |
| TelperionAI/EXL3-4.0bpw publisher's own comparison build [L33]publisher comparison build | TelperionAI NVFP4 (vLLM) | EXL3 | 24.70 | download | — | 93.62% | — | — | — | vs BF16 | L33 @ main |
| TelperionAI/EXL3-4.0bpw sibling checkpoint in series [L30]publisher comparison build | EXL3 6.5 bpw | EXL3 | 23.00 | download | — | 97.10% | — | — | — | vs BF16 | L30 @ main |
| TelperionAI/EXL3-4.0bpw sibling checkpoint in series [L31]publisher comparison build | EXL3 5.5 bpw | EXL3 | 20.00 | download | — | 96.12% | — | — | — | vs BF16 | L31 @ main |
| TelperionAI/EXL3-4.0bpw [L35]publisher comparison build | EXL3 4.0 bpw, no AWQ pre-pass | EXL3 | 16.00 | download | — | 93.17% | — | — | — | vs BF16 | L35 @ main |
| TelperionAI/EXL3-4.0bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38]publisher comparison build | TelperionAI FP8-block-AWQ | EXL3 | 30.00 | download | — | 96.22% | — | — | — | vs BF16 | L44 @ main |
| TelperionAI/EXL3-4.0bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38]publisher comparison build | TelperionAI INT4-AWQ-GPTQ | EXL3 | 25.10 | download | — | 96.30% | — | — | — | vs BF16 | L45 @ main |
| TelperionAI/EXL3-4.0bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38]publisher comparison build | TelperionAI NVFP4 (mixed) | EXL3 | 24.70 | download | — | 93.62% | — | — | — | vs BF16 | L51 @ main |
| TelperionAI/EXL3-4.0bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38]publisher comparison build | TelperionAI NVFP4 (full) | EXL3 | 24.70 | download | — | 93.05% | — | — | — | vs BF16 | L52 @ main |
| TelperionAI/EXL3-4.0bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38]publisher comparison build | TelperionAI EXL3 6.5bpw (v1) | EXL3 | 23.00 | download | — | 97.10% | — | — | — | vs BF16 | L50 @ main |
| TelperionAI/EXL3-4.0bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38]publisher comparison build | TelperionAI EXL3 5.5bpw-v2 | EXL3 | 20.00 | download | — | 96.26% | — | — | — | vs BF16 | L46 @ main |
| TelperionAI/EXL3-4.0bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38]publisher comparison build | TelperionAI EXL3 5.5bpw (v1) | EXL3 | 20.00 | download | — | 96.12% | — | — | — | vs BF16 | L49 @ main |
| TelperionAI/EXL3-4.0bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38]publisher comparison build | TelperionAI EXL3 4.0bpw (v1) | EXL3 | 16.00 | download | — | 93.82% | — | — | — | vs BF16 | L48 @ main |
| TelperionAI/EXL3-5.5bpw publisher's own comparison build [L29]publisher comparison build | TelperionAI INT4-AWQ-GPTQ (vLLM) | EXL3 | 25.10 | download | — | 96.30% | — | — | — | vs BF16 | L29 @ main |
| TelperionAI/EXL3-5.5bpw publisher's own comparison build [L33]publisher comparison build | TelperionAI NVFP4 (vLLM) | EXL3 | 24.70 | download | — | 93.62% | — | — | — | vs BF16 | L33 @ main |
| TelperionAI/EXL3-5.5bpw sibling checkpoint in series [L30]publisher comparison build | EXL3 6.5 bpw | EXL3 | 23.00 | download | — | 97.10% | — | — | — | vs BF16 | L30 @ main |
| TelperionAI/EXL3-5.5bpw sibling checkpoint in series [L34]publisher comparison build | EXL3 4.0 bpw | EXL3 | 16.00 | download | — | 93.82% | — | — | — | vs BF16 | L34 @ main |
| TelperionAI/EXL3-5.5bpw [L35]publisher comparison build | EXL3 4.0 bpw, no AWQ pre-pass | EXL3 | 16.00 | download | — | 93.17% | — | — | — | vs BF16 | L35 @ main |
| TelperionAI/EXL3-5.5bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38]publisher comparison build | TelperionAI FP8-block-AWQ | EXL3 | 30.00 | download | — | 96.22% | — | — | — | vs BF16 | L44 @ main |
| TelperionAI/EXL3-5.5bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38]publisher comparison build | TelperionAI INT4-AWQ-GPTQ | EXL3 | 25.10 | download | — | 96.30% | — | — | — | vs BF16 | L45 @ main |
| TelperionAI/EXL3-5.5bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38]publisher comparison build | TelperionAI NVFP4 (mixed) | EXL3 | 24.70 | download | — | 93.62% | — | — | — | vs BF16 | L51 @ main |
| TelperionAI/EXL3-5.5bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38]publisher comparison build | TelperionAI NVFP4 (full) | EXL3 | 24.70 | download | — | 93.05% | — | — | — | vs BF16 | L52 @ main |
| TelperionAI/EXL3-5.5bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38]publisher comparison build | TelperionAI EXL3 6.5bpw (v1) | EXL3 | 23.00 | download | — | 97.10% | — | — | — | vs BF16 | L50 @ main |
| TelperionAI/EXL3-5.5bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38]publisher comparison build | TelperionAI EXL3 5.5bpw (v1) | EXL3 | 20.00 | download | — | 96.12% | — | — | — | vs BF16 | L49 @ main |
| TelperionAI/EXL3-5.5bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38]publisher comparison build | TelperionAI EXL3 4.0bpw-v2 | EXL3 | 16.00 | download | — | 95.16% | — | — | — | vs BF16 | L46 @ main |
| TelperionAI/EXL3-5.5bpw-v2 disagreement vs BF16 base by confidence bucket; 'confident' is the pre-registered decider [L35-38]publisher comparison build | TelperionAI EXL3 4.0bpw (v1) | EXL3 | 16.00 | download | — | 93.82% | — | — | — | vs BF16 | L48 @ main |
| TelperionAI/EXL3-6.5bpw publisher's own comparison build [L29]publisher comparison build | TelperionAI INT4-AWQ-GPTQ (vLLM) | EXL3 | 25.10 | download | — | 96.30% | — | — | — | vs BF16 | L29 @ main |
| TelperionAI/EXL3-6.5bpw publisher's own comparison build [L33]publisher comparison build | TelperionAI NVFP4 (vLLM) | EXL3 | 24.70 | download | — | 93.62% | — | — | — | vs BF16 | L33 @ main |
| TelperionAI/EXL3-6.5bpw sibling checkpoint in series [L31]publisher comparison build | EXL3 5.5 bpw | EXL3 | 20.00 | download | — | 96.12% | — | — | — | vs BF16 | L31 @ main |
| TelperionAI/EXL3-6.5bpw sibling checkpoint in series [L34]publisher comparison build | EXL3 4.0 bpw | EXL3 | 16.00 | download | — | 93.82% | — | — | — | vs BF16 | L34 @ main |
| TelperionAI/EXL3-6.5bpw [L35]publisher comparison build | EXL3 4.0 bpw, no AWQ pre-pass | EXL3 | 16.00 | download | — | 93.17% | — | — | — | vs BF16 | L35 @ main |
| TelperionAI/FP8-block-AWQ [L31]publisher comparison build | TelperionAI INT4-AWQ-GPTQ | AWQ | 25.10 | unknown | — | 96.30% | — | — | — | vs BF16 | L31 @ main |
| TelperionAI/FP8-block-AWQ [L35]publisher comparison build | TelperionAI NVFP4 | AWQ | 24.70 | unknown | — | 93.62% | — | — | — | vs BF16 | L35 @ main |
| TelperionAI/FP8-block-AWQ [L32]publisher comparison build | TelperionAI EXL3 6.5 bpw + AWQ | AWQ | 23.00 | unknown | — | 97.10% | — | — | — | vs BF16 | L32 @ main |
| TelperionAI/FP8-block-AWQ [L33]publisher comparison build | TelperionAI EXL3 5.5 bpw + AWQ | AWQ | 20.00 | unknown | — | 96.12% | — | — | — | vs BF16 | L33 @ main |
| TelperionAI/INT4-AWQ-GPTQ repo not named on cardpublisher comparison build | our NVFP4 build (AWQ+AutoRound) | AWQ | 24.70 | download | — | 93.38% | — | — | — | vs BF16 | L57 @ main |
| TelperionAI/INT4-AWQ-GPTQ-gdn4 [L60]publisher comparison build | our GDN 8-bit sibling (Qwen3.8-27B-INT4-AWQ-GPTQ) | AWQ | 25.10 | download | — | 96.30% | — | — | near-tie disagree 22.29%; moderate disagree 3.52%; confident disagree 0.93%; certain disagree 0.09%; divmed 48 | vs BF16 | L60 @ main |
| TelperionAI/NVFP4-AWQ-AutoRound-allfp4 [L58]publisher comparison build | sibling with FP8 MLP 56-63 | NVFP4 | 24.70 | download | — | 93.38% | — | — | — | vs BF16 | L58 @ main |
| ThakiCloud/NVFP4-FP8ATTN GSM8K delta vs bf16 -5.50pp, p 0.056 [L91]publisher comparison build | our pure NVFP4 (MLP only, GPTQ-txt) | NVFP4 | 30.14 | download | — | — | — | — | MMMU-val MC 0.8922 | vs BF16 | L44 @ main |
| ThakiCloud/W4A16-AWQ same model, same target size [L13-L14]publisher comparison build | GPTQ sibling | AWQ | — | — | — | — | — | GSM8K_en_strict 0.5133; GSM8K_en_flexible 0.5641; GSM8K_ko_strict 0.5686; GSM8K_ko_flexible 0.6505 | vs BF16 | L26 @ main | |
| ThakiCloud/W4A16-AWQ unquantized reference [L25]publisher comparison build | bf16 | AWQ | — | — | — | — | — | GSM8K_en_strict 0.5542; GSM8K_en_flexible 0.5550; GSM8K_ko_strict 0.6149; GSM8K_ko_flexible 0.6315 | vs BF16 | L25 @ main | |
| thederby/oQ4e-fp16-mtp reference model [L34]publisher comparison build | original bf16 | other | 56.00 | download | 0 | 100% | 10.20 | — | — | vs BF16 | L34 @ main |
| thederby/oQ4e-fp16-mtp [L33]publisher comparison build | oQ8e fp16 | other | 30.90 | download | 0.0007 | 98.4% | 10.20 | — | — | vs BF16 | L33 @ main |
| thederby/oQ4e-fp16-mtp [L32]publisher comparison build | oQ8e bf16 | other | 30.00 | download | 0.0010 | 98.6% | 10.21 | — | — | vs BF16 | L32 @ main |
| thederby/oQ4e-fp16-mtp [L31]publisher comparison build | oQ6e fp16 | other | 24.60 | download | 0.0028 | 97.1% | 10.22 | — | — | vs BF16 | L31 @ main |
| thederby/oQ4e-fp16-mtp [L30]publisher comparison build | oQ6e bf16 | other | 23.70 | download | 0.0033 | 97.2% | 10.23 | — | — | vs BF16 | L30 @ main |
| thederby/oQ4e-fp16-mtp [L29]publisher comparison build | oQ5e fp16 | other | 21.20 | download | 0.0103 | 94.9% | 10.28 | — | — | vs BF16 | L29 @ main |
| thederby/oQ4e-fp16-mtp [L28]publisher comparison build | oQ5e bf16 | other | 20.30 | download | 0.0106 | 94.9% | 10.28 | — | — | vs BF16 | L28 @ main |
| thederby/oQ4e-fp16-mtp [L26]publisher comparison build | oQ4e bf16 | other | 17.00 | download | 0.0358 | 90.6% | 10.47 | — | — | vs BF16 | L26 @ main |
| thederby/oQ4e-mtp reference row [L34]publisher comparison build | original bf16 | other | 56.00 | download | 0 | 100% | 10.20 | — | — | vs BF16 | L34 @ main |
| thederby/oQ4e-mtp [L33]publisher comparison build | oQ8e fp16 | other | 30.90 | download | 0.0007 | 98.4% | 10.20 | — | — | vs BF16 | L33 @ main |
| thederby/oQ4e-mtp [L32]publisher comparison build | oQ8e bf16 | other | 30.00 | download | 0.0010 | 98.6% | 10.21 | — | — | vs BF16 | L32 @ main |
| thederby/oQ4e-mtp [L31]publisher comparison build | oQ6e fp16 | other | 24.60 | download | 0.0028 | 97.1% | 10.22 | — | — | vs BF16 | L31 @ main |
| thederby/oQ4e-mtp [L30]publisher comparison build | oQ6e bf16 | other | 23.70 | download | 0.0033 | 97.2% | 10.23 | — | — | vs BF16 | L30 @ main |
| thederby/oQ4e-mtp [L29]publisher comparison build | oQ5e fp16 | other | 21.20 | download | 0.0103 | 94.9% | 10.28 | — | — | vs BF16 | L29 @ main |
| thederby/oQ4e-mtp [L28]publisher comparison build | oQ5e bf16 | other | 20.30 | download | 0.0106 | 94.9% | 10.28 | — | — | vs BF16 | L28 @ main |
| thederby/oQ4e-mtp [L27]; vision tower float32 [L39]publisher comparison build | oQ4e fp16 | other | 17.90 | download | 0.0351 | 90.5% | 10.47 | — | — | vs BF16 | L27 @ main |
| thederby/oQ5e-fp16-mtp reference row [L34]publisher comparison build | original bf16 | other | 56.00 | download | — | — | 10.20 | — | — | vs BF16 | L34 @ main |
| thederby/oQ5e-fp16-mtp [L33]publisher comparison build | oQ8e fp16 | other | 30.90 | download | 0.0007 | 98.4% | 10.20 | — | — | vs BF16 | L33 @ main |
| thederby/oQ5e-fp16-mtp [L32]publisher comparison build | oQ8e bf16 | other | 30.00 | download | 0.0010 | 98.6% | 10.21 | — | — | vs BF16 | L32 @ main |
| thederby/oQ5e-fp16-mtp [L31]publisher comparison build | oQ6e fp16 | other | 24.60 | download | 0.0028 | 97.1% | 10.22 | — | — | vs BF16 | L31 @ main |
| thederby/oQ5e-fp16-mtp [L30]publisher comparison build | oQ6e bf16 | other | 23.70 | download | 0.0033 | 97.2% | 10.23 | — | — | vs BF16 | L30 @ main |
| thederby/oQ5e-fp16-mtp [L28]publisher comparison build | oQ5e bf16 | other | 20.30 | download | 0.0106 | 94.9% | 10.28 | — | — | vs BF16 | L28 @ main |
| thederby/oQ5e-fp16-mtp [L27]publisher comparison build | oQ4e fp16 | other | 17.90 | download | 0.0351 | 90.5% | 10.47 | — | — | vs BF16 | L27 @ main |
| thederby/oQ5e-fp16-mtp [L26]publisher comparison build | oQ4e bf16 | other | 17.00 | download | 0.0358 | 90.6% | 10.47 | — | — | vs BF16 | L26 @ main |
| thederby/oQ5e-mtp reference row [L34]publisher comparison build | original bf16 | other | 56.00 | download | 0 | 100% | 10.20 | — | — | vs BF16 | L34 @ main |
| thederby/oQ5e-mtp publisher comparison build | oQ8e fp16 | other | 30.90 | download | 0.0007 | 98.4% | 10.20 | — | — | vs BF16 | L33 @ main |
| thederby/oQ5e-mtp publisher comparison build | oQ8e bf16 | other | 30.00 | download | 0.0010 | 98.6% | 10.21 | — | — | vs BF16 | L32 @ main |
| thederby/oQ5e-mtp publisher comparison build | oQ6e fp16 | other | 24.60 | download | 0.0028 | 97.1% | 10.22 | — | — | vs BF16 | L31 @ main |
| thederby/oQ5e-mtp publisher comparison build | oQ6e bf16 | other | 23.70 | download | 0.0033 | 97.2% | 10.23 | — | — | vs BF16 | L30 @ main |
| thederby/oQ5e-mtp ~0.9 GB larger, vision tower in float32 [L38-39]publisher comparison build | oQ5e fp16 | other | 21.20 | download | 0.0103 | 94.9% | 10.28 | — | — | vs BF16 | L29 @ main |
| thederby/oQ5e-mtp publisher comparison build | oQ4e fp16 | other | 17.90 | download | 0.0351 | 90.5% | 10.47 | — | — | vs BF16 | L27 @ main |
| thederby/oQ5e-mtp publisher comparison build | oQ4e bf16 | other | 17.00 | download | 0.0358 | 90.6% | 10.47 | — | — | vs BF16 | L26 @ main |
| thederby/oQ6e-fp16-mtp reference row [L34]publisher comparison build | original bf16 | other | 56.00 | download | 0 | 100% | 10.20 | — | — | vs BF16 | L34 @ main |
| thederby/oQ6e-fp16-mtp publisher comparison build | oQ8e fp16 | other | 30.90 | download | 0.0007 | 98.4% | 10.20 | — | — | vs BF16 | L33 @ main |
| thederby/oQ6e-fp16-mtp publisher comparison build | oQ8e bf16 | other | 30.00 | download | 0.0010 | 98.6% | 10.21 | — | — | vs BF16 | L32 @ main |
| thederby/oQ6e-fp16-mtp publisher comparison build | oQ6e bf16 | other | 23.70 | download | 0.0033 | 97.2% | 10.23 | — | — | vs BF16 | L30 @ main |
| thederby/oQ6e-fp16-mtp publisher comparison build | oQ5e fp16 | other | 21.20 | download | 0.0103 | 94.9% | 10.28 | — | — | vs BF16 | L29 @ main |
| thederby/oQ6e-fp16-mtp publisher comparison build | oQ5e bf16 | other | 20.30 | download | 0.0106 | 94.9% | 10.28 | — | — | vs BF16 | L28 @ main |
| thederby/oQ6e-fp16-mtp fp16 variants ~0.9 GB larger, vision tower kept float32 [L39]publisher comparison build | oQ4e fp16 | other | 17.90 | download | 0.0351 | 90.5% | 10.47 | — | — | vs BF16 | L27 @ main |
| thederby/oQ6e-fp16-mtp publisher comparison build | oQ4e bf16 | other | 17.00 | download | 0.0358 | 90.6% | 10.47 | — | — | vs BF16 | L26 @ main |
| thederby/oQ6e-mtp reference row [L34]publisher comparison build | original bf16 | other | 56.00 | download | — | — | 10.20 | — | — | vs BF16 | L34 @ main |
| thederby/oQ6e-mtp publisher comparison build | oQ8e fp16 | other | 30.90 | download | 0.0007 | 98.4% | 10.20 | — | — | vs BF16 | L33 @ main |
| thederby/oQ6e-mtp publisher comparison build | oQ8e bf16 | other | 30.00 | download | 0.0010 | 98.6% | 10.21 | — | — | vs BF16 | L32 @ main |
| thederby/oQ6e-mtp vision tower kept float32, about 0.9 GB larger [L38-39]publisher comparison build | oQ6e fp16 | other | 24.60 | download | 0.0028 | 97.1% | 10.22 | — | — | vs BF16 | L31 @ main |
| thederby/oQ6e-mtp publisher comparison build | oQ5e fp16 | other | 21.20 | download | 0.0103 | 94.9% | 10.28 | — | — | vs BF16 | L29 @ main |
| thederby/oQ6e-mtp publisher comparison build | oQ5e bf16 | other | 20.30 | download | 0.0106 | 94.9% | 10.28 | — | — | vs BF16 | L28 @ main |
| thederby/oQ6e-mtp publisher comparison build | oQ4e fp16 | other | 17.90 | download | 0.0351 | 90.5% | 10.47 | — | — | vs BF16 | L27 @ main |
| thederby/oQ6e-mtp publisher comparison build | oQ4e bf16 | other | 17.00 | download | 0.0358 | 90.6% | 10.47 | — | — | vs BF16 | L26 @ main |
| thederby/oQ8e-fp16-mtp reference row [L34]publisher comparison build | original bf16 | other | 56.00 | download | 0 | 100% | 10.20 | — | — | vs BF16 | L34 @ main |
| thederby/oQ8e-fp16-mtp [L32]publisher comparison build | oQ8e bf16 | other | 30.00 | download | 0.0010 | 98.6% | 10.21 | — | — | vs BF16 | L32 @ main |
| thederby/oQ8e-fp16-mtp [L31]publisher comparison build | oQ6e fp16 | other | 24.60 | download | 0.0028 | 97.1% | 10.22 | — | — | vs BF16 | L31 @ main |
| thederby/oQ8e-fp16-mtp [L30]publisher comparison build | oQ6e bf16 | other | 23.70 | download | 0.0033 | 97.2% | 10.23 | — | — | vs BF16 | L30 @ main |
| thederby/oQ8e-fp16-mtp [L29]publisher comparison build | oQ5e fp16 | other | 21.20 | download | 0.0103 | 94.9% | 10.28 | — | — | vs BF16 | L29 @ main |
| thederby/oQ8e-fp16-mtp [L28]publisher comparison build | oQ5e bf16 | other | 20.30 | download | 0.0106 | 94.9% | 10.28 | — | — | vs BF16 | L28 @ main |
| thederby/oQ8e-fp16-mtp [L27]publisher comparison build | oQ4e fp16 | other | 17.90 | download | 0.0351 | 90.5% | 10.47 | — | — | vs BF16 | L27 @ main |
| thederby/oQ8e-fp16-mtp [L26]publisher comparison build | oQ4e bf16 | other | 17.00 | download | 0.0358 | 90.6% | 10.47 | — | — | vs BF16 | L26 @ main |
| thederby/oQ8e-mtp reference rowpublisher comparison build | original bf16 | other | 56.00 | download | 0 | 100% | 10.20 | — | — | vs BF16 | L34 @ main |
| thederby/oQ8e-mtp fp16 variant ~0.9 GB larger, vision tower float32 [L39]publisher comparison build | oQ8e fp16 | other | 30.90 | download | 0.0007 | 98.4% | 10.20 | — | — | vs BF16 | L33 @ main |
| thederby/oQ8e-mtp publisher comparison build | oQ6e fp16 | other | 24.60 | download | 0.0028 | 97.1% | 10.22 | — | — | vs BF16 | L31 @ main |
| thederby/oQ8e-mtp publisher comparison build | oQ6e bf16 | other | 23.70 | download | 0.0033 | 97.2% | 10.23 | — | — | vs BF16 | L30 @ main |
| thederby/oQ8e-mtp publisher comparison build | oQ5e fp16 | other | 21.20 | download | 0.0103 | 94.9% | 10.28 | — | — | vs BF16 | L29 @ main |
| thederby/oQ8e-mtp publisher comparison build | oQ5e bf16 | other | 20.30 | download | 0.0106 | 94.9% | 10.28 | — | — | vs BF16 | L28 @ main |
| thederby/oQ8e-mtp publisher comparison build | oQ4e fp16 | other | 17.90 | download | 0.0351 | 90.5% | 10.47 | — | — | vs BF16 | L27 @ main |
| thederby/oQ8e-mtp publisher comparison build | oQ4e bf16 | other | 17.00 | download | 0.0358 | 90.6% | 10.47 | — | — | vs BF16 | L26 @ main |
| TheDrainFlorist/VQ-4.5bpw superseded prose-only table [L123]publisher comparison build | affine q6 (ours) | other | 22.77 | download | — | — | — | — | kl_mnats:prose_2048tok 3.71; top1:prose_2048tok 96.8%; ppl:prose_2048tok 5.260 | vs BF16 | L123 @ main |
| TheDrainFlorist/VQ-4.5bpw superseded prose-only table [L120]publisher comparison build | affine q3 (ours) | other | 12.69 | download | — | — | — | — | kl_mnats:prose_2048tok 187.8; top1:prose_2048tok 79.5%; ppl:prose_2048tok 5.832 | vs BF16 | L120 @ main |
| TheDrainFlorist/VQ-4.5bpw superseded prose-only table [L119]publisher comparison build | affine q2 (ours) | other | 9.3 | download | — | — | — | — | kl_mnats:prose_2048tok 1426.9; top1:prose_2048tok 46.1%; ppl:prose_2048tok 16.435 | vs BF16 | L119 @ main |
| TheDrainFlorist/VQ-4.8bpw [L124]publisher comparison build | affine q8 (ours) (superseded table) | other | 29.51 | download | 1.25 mnats | 98.5% | 5.241 | — | — | vs BF16 | L124 @ main |
| TheDrainFlorist/VQ-4.8bpw [L123]publisher comparison build | affine q6 (ours) (superseded table) | other | 22.77 | download | 3.71 mnats | 96.8% | 5.260 | — | — | vs BF16 | L123 @ main |
| TheDrainFlorist/VQ-4.8bpw [L121]publisher comparison build | affine q4 (ours) (superseded table) | other | 16.05 | download | 45.8 mnats | 89.8% | 5.206 | — | — | vs BF16 | L121 @ main |
| TheDrainFlorist/VQ-4.8bpw [L120]publisher comparison build | affine q3 (ours) (superseded table) | other | 12.69 | download | 187.8 mnats | 79.5% | 5.832 | — | — | vs BF16 | L120 @ main |
| TheDrainFlorist/VQ-4.8bpw [L119]publisher comparison build | affine q2 (ours) (superseded table) | other | 9.3 | download | 1426.9 mnats | 46.1% | 16.435 | — | — | vs BF16 | L119 @ main |
| timteh673/Opus-Abliterix-Reasoning-MLX-8bit reference row; KL 0 by definition [L52]publisher comparison build | control-bf16 | MLX | — | 0.000000 | — | — | — | Capability macro 17.6859%; Full code 16; HumanEval 7.9268%; Long-form pass 54.1667%; MMMU30 9; Harmful hard refusal 43.2% | vs BF16 | L43 @ main | |
| ToBeStyled/ColdFusion-GAIN-Blackwell-DFlash2-Ultra-V1.0 [L113]publisher comparison build | Q4_K_M of the same finetune + DFlash2 | other | 18.50 | download | — | — | 2.7325 ± 0.033 | — | — | other or no reference | L113 @ main |
| ToBeStyled/ColdFusion-GAIN-Blackwell-DFlash2-Ultra-V1.0 size from Notes text [L128]publisher comparison build | Fable NVFP4 + imatrix scale search | other | 16.00 | download | — | — | 2.8016 | — | — | other or no reference | L128 @ main |
| ToBeStyled/ColdFusion-GAIN-Blackwell-DFlash2-Ultra-V1.0 Fable variants requantized by hand; relation to this repo's finetune not stated [L123-125]publisher comparison build | Fable Q4_K_M (the bar) | other | — | — | — | 2.7660 | — | — | other or no reference | L127 @ main | |
| ToBeStyled/ColdFusion-GAIN-Blackwell-DFlash2-Ultra-V1.0 [L129]publisher comparison build | Fable NVFP4 legacy fixed amax/6 | other | — | — | — | 2.8231 | — | — | other or no reference | L129 @ main | |
| ToBeStyled/ColdFusion-GAIN-Blackwell-DFlash2-Ultra-V1.0 22 GB per notes [L130]publisher comparison build | Fable NVFP4 + FFN/attention in Q6_K | other | — | — | — | 2.7573 | — | — | other or no reference | L130 @ main | |
| trevest/Tetris-GGUF Reference row, +0.13% vs BF16 [L31]publisher comparison build | Q8_0 reference | GGUF | 29.04 | download | — | — | 6.9887 | — | — | vs a Q8_0 | L26 @ main |
| trohrbaugh/heretic-ara measured by ayushcluster00KL vs base of source abliteration [L129] | trohrbaugh -ara (source) | — | 0.0535 | — | — | — | — | other or no reference | L129 @ main | ||
| trohrbaugh/heretic-ara measured by jaromerabliteration source model, KL vs base; refusals 3/100 [L129] | trohrbaugh -ara (source) | — | 0.0535 | — | — | — | — | other or no reference | L129 @ main | ||
| trohrbaugh/heretic-ara measured by slevinwKL vs base model (abliteration), not quantization [L131] | trohrbaugh -ara (source) | — | 0.0535 | — | — | — | — | other or no reference | L131 @ main | ||
| turboderp/exl3 measured by malaiwah21.37 GiB file / 17.05 body [L453] | turboderp/Qwen3.8-27B-exl3 6.00bpw | EXL3 | 22.95 | download | 0.001583 | 98.28 % | — | — | — | vs BF16 | L453 @ main |
| turboderp/exl3 measured by malaiwahfile size; body 17.05 GiB c78.1 [L462] | turboderp/Qwen3.8-27B-exl3 6.00bpw @ d32ba0bb | EXL3 | 22.95 | download | 0.001583 | 98.28 % | — | — | — | vs BF16 | L462 @ main |
| turboderp/exl3 measured by malaiwahfile size 21.37 GiB, body 17.05 [L493] | turboderp/Qwen3.8-27B-exl3 6.00bpw shard0 | EXL3 | 22.95 | download | 0.001583 | 98.28 % | — | — | — | other or no reference | L493 @ main |
| turboderp/exl3 measured by malaiwah18.53 GiB file / 14.22 body [L458] | turboderp/Qwen3.8-27B-exl3 5.00bpw | EXL3 | 19.90 | download | 0.004005 | 97.37 % | — | — | — | vs BF16 | L458 @ main |
| turboderp/exl3 measured by malaiwahfile size; body 14.22 GiB c97.1 [L467] | turboderp/Qwen3.8-27B-exl3 5.00bpw @ a35e75a7 | EXL3 | 19.90 | download | 0.004005 | 97.37 % | — | — | — | vs BF16 | L467 @ main |
| turboderp/exl3 measured by malaiwahfile size 18.53 GiB, body 14.22 [L498] | turboderp/Qwen3.8-27B-exl3 5.00bpw shard0 | EXL3 | 19.90 | download | 0.004005 | 97.37 % | — | — | — | other or no reference | L498 @ main |
| Twu31/AWQ-INT4-MTP-LowLatency [L147]publisher comparison build | INT4 body, BF16 lm_head | AWQ | — | — | — | — | — | NLL 0.327 | vs BF16 | L147 @ main | |
| Twu31/AWQ-INT4-MTP-LowLatency [L149]; 17.6 GB in prose [L248]publisher comparison build | INT4 + INT4 lm_head (lm-head-int4) | AWQ | — | — | — | — | — | NLL 0.326 | vs BF16 | L149 @ main | |
| Twu31/AWQ-INT4-MTP-LowLatency [L150]; 19.0 GB in prose [L248]publisher comparison build | MTP head also INT4 (mtp-int4-autoround) | AWQ | — | — | — | — | — | NLL 0.324 | vs BF16 | L150 @ main | |
| UgurI/Turkish-CPT-GGUF Comparison column in the card; file/precision of base run not stated [L36]publisher comparison build | Base Model (Qwen3.8-27B) | GGUF | — | — | — | — | — | ppl:turkish_validation 7.5838; ppl:turkish_test 8.0534; TurkishMMLU 60.00%; Belebele TR 83.33%; Belebele EN 96.67% | other or no reference | L38 @ main | |
| unsloth/GGUF measured by magiccodingmanRebuilt by MagicQuant under controlled conditions with its own imatrix, not the original artifact [L72] | UD-Unsloth-UD-Q6_K_XL | GGUF | 25.33 | download | 0.001238 | — | — | — | — | reference not stated | L39 @ main |
| unsloth/GGUF measured by magiccodingmanRebuilt under MagicQuant flow [L72] | UD-Unsloth-UD-Q6_K_L | GGUF | 24.23 | download | 0.001439 | — | — | — | — | reference not stated | L40 @ main |
| unsloth/GGUF measured by magiccodingmanRebuilt under MagicQuant flow [L72] | UD-Unsloth-UD-Q4_K_M | GGUF | 16.50 | download | 0.011205 | — | — | — | — | reference not stated | L47 @ main |
| unsloth/GGUF measured by oshitekuNo SWE-bench value in this column | Qwen3.8 UD-Q4_K_M | GGUF | 16.50 | download | 0.014 | 97.0% | — | — | — | vs BF16 | L70 @ main |
| unsloth/GGUF measured by fraQtlrepo id completed from truncated link text 'unsloth/Qwen3.8-2...' [L35] | unsloth canonical Q4_K_M | GGUF | 16.46 | download | — | — | — | — | kld:code/math 0.22493; kld:general 0.28209; GSM8K 94.0% (188/200); MATH-500 88.2% (441/500) | vs a Q8_0 | L35 @ main |
| unsloth/GGUF measured by akopytko[L62,L70] | unsloth Q4_0 | GGUF | 16.05 | download | 0.028317 +/- 0.000330 | 92.698 +/- 0.115% | — | — | — | vs BF16 | L62 @ main |
| unsloth/GGUF measured by magiccodingmanRebuilt under MagicQuant flow [L72] | UD-Unsloth-UD-Q4_K_S | GGUF | 15.39 | download | 0.014502 | — | — | — | — | reference not stated | L49 @ main |
| unsloth/GGUF measured by magiccodingmanRebuilt under MagicQuant flow [L72] | UD-Unsloth-UD-IQ4_XS | GGUF | 14.29 | download | 0.020127 | — | — | — | — | reference not stated | L50 @ main |
| unsloth/GGUF measured by Luceboxmeasured by card author as comparison [L21] | Unsloth UD-IQ4_XS | GGUF | 14.25 | download | 0.018 | 94.1% | — | +0.7% | — | vs a Q8_0 | L23 @ main |
| unsloth/GGUF measured by jdominick05repo id inferred from the recipe source link [L104]; table row only says Unsloth UD-IQ4_XS [L46] | Unsloth UD-IQ4_XS | GGUF | 14.25 | download | 0.0273 ± 0.0017 | 94.74% | — | +1.0% | — | vs a Q8_0 | L46 @ main |
| unsloth/GGUF measured by wiklifrepo inferred [L391] | Unsloth UD-IQ4_XS | GGUF | 14.25 | download | 0.0187 kld:prose | — | — | — | kld:code 0.0138 | vs a Q8_0 | L112 @ main |
| unsloth/GGUF measured by magiccodingmanRebuilt under MagicQuant flow [L72] | UD-Unsloth-UD-Q3_K_XL | GGUF | 13.18 | download | 0.030770 | — | — | — | — | reference not stated | L51 @ main |
| unsloth/GGUF measured by wiklifrepo inferred [L391] | Unsloth UD-Q3_K_XL | GGUF | 13.15 | download | 0.0287 kld:prose | — | — | — | kld:code 0.0228 | vs a Q8_0 | L113 @ main |
| unsloth/GGUF measured by magiccodingmanRebuilt under MagicQuant flow [L72] | UD-Unsloth-UD-IQ3_S | GGUF | 12.08 | download | 0.046499 | — | — | — | — | reference not stated | L52 @ main |
| unsloth/GGUF measured by wiklifrepo inferred [L391] | Unsloth UD-IQ3_S | GGUF | 12.04 | download | 0.0507 kld:prose | — | — | — | kld:code 0.0330 | vs a Q8_0 | L114 @ main |
| unsloth/GGUF measured by magiccodingmanRebuilt under MagicQuant flow [L72] | UD-Unsloth-UD-IQ3_XXS | GGUF | 10.97 | download | 0.070084 | — | — | — | — | reference not stated | L54 @ main |
| unsloth/GGUF measured by wiklifrepo inferred from reference/credits lines [L55,L391] | Unsloth UD-IQ3_XXS | GGUF | 10.93 | download | 0.0781 kld:prose | — | — | — | GPQA Diamond 87.9 ± 4.5; IFBench_loose 83.0; IFBench_strict 80.3; LiveCodeBench v6 87.0 ± 5.8; kld:code 0.0487 | vs a Q8_0 | L67 @ main |
| unsloth/GGUF measured by magiccodingmanRebuilt under MagicQuant flow [L72] | UD-Unsloth-UD-Q2_K_XL | GGUF | 9.9 | download | 0.105510 | — | — | — | — | reference not stated | L56 @ main |
| unsloth/GGUF measured by wiklifrepo inferred [L391] | Unsloth UD-Q2_K_XL | GGUF | 9.8 | download | 0.1061 kld:prose | — | — | — | kld:code 0.0741 | vs a Q8_0 | L119 @ main |
| unsloth/GGUF measured by magiccodingmanRebuilt under MagicQuant flow [L72] | UD-Unsloth-UD-IQ2_XXS | GGUF | 9.1 | download | 0.182364 | — | — | — | — | reference not stated | L57 @ main |
| unsloth/GGUF measured by Rabbit-Hole-Airepo inferred from card links [L288]; 2.17x this cook KLD [L186] | Unsloth UD-Q6_K | GGUF | — | — | — | — | — | kld:corpus A 0.003486; top1:corpus A 97.173% | vs BF16 | L186 @ main | |
| unsloth/GGUF measured by Rabbit-Hole-Airepo inferred from card links [L288]; 3.22x this cook KLD [L187] | Unsloth UD-Q5_K_XL | GGUF | — | — | — | — | — | kld:corpus A 0.005169; top1:corpus A 96.756% | vs BF16 | L187 @ main | |
| unsloth/GGUF measured by Rabbit-Hole-Airepo inferred from card links [L288]; 4.19x this cook KLD [L188] | Unsloth UD-Q5_K_M | GGUF | — | — | — | — | — | kld:corpus A 0.006734; top1:corpus A 96.202% | vs BF16 | L188 @ main | |
| unsloth/GGUF measured by ToBeStyled[L49,L67-68] | Base Q4_K_M (UD-Q4_K_M, no speculation) | GGUF | — | — | — | — | — | HumanEval-50 pass@1 84.0%; GSM8K-50 94.0% | other or no reference | L67 @ main | |
| unsloth/GGUF quoted by acyildirimer, not measured[L82] | unsloth GGUF Q8_0 | NVFP4 | — | — | — | — | — | MMLU 85.2; ARC-C 96.8; HellaSwag 94.4; GSM8K 97.4; HumanEval 94.5; 4-task mean 93.5 | vs BF16 | L82 @ main | |
| unsloth/GGUF quoted by acyildirimer, not measured[L83] | unsloth GGUF Q6_K | NVFP4 | — | — | — | — | — | MMLU 85.3; ARC-C 96.7; HellaSwag 94.3; GSM8K 97.5; HumanEval 94.5; 4-task mean 93.5 | vs BF16 | L83 @ main | |
| unsloth/GGUF quoted by acyildirimer, not measured[L84] | unsloth GGUF UD-Q4_K_XL | NVFP4 | — | — | — | — | — | MMLU 85.1; ARC-C 96.6; HellaSwag 94.4; GSM8K 97.3; HumanEval 93.9; 4-task mean 93.4 | vs BF16 | L84 @ main | |
| unsloth/GGUF quoted by acyildirimer, not measured[L85] | unsloth GGUF Q4_K_M | NVFP4 | — | — | — | — | — | MMLU 85.0; ARC-C 96.8; HellaSwag 94.3; GSM8K 97.1; HumanEval 92.7; 4-task mean 93.3 | vs BF16 | L85 @ main | |
| unsloth/GGUF measured by akopytkorepo inferred from the unsloth link at [L62]; [L71] | unsloth Q6_K | GGUF | — | 0.002857 +/- 0.000051 | 97.506 +/- 0.069% | — | — | — | vs BF16 | L71 @ main | |
| unsloth/GGUF measured by jcbtcComparison files from Unsloth Dynamic 3.0, revision 4ca720788d1e [L121]; run at 65,536 context [L115] | Unsloth Dynamic Q4 | GGUF | — | — | — | — | — | HumanEval 158/164 (96.34%); HumanEval+ 148/164 (90.24%) | reference not stated | L105 @ main | |
| unsloth/GGUF measured by jcbtcComparison files from Unsloth Dynamic 3.0 [L121]; run at 65,536 context [L115] | Unsloth Dynamic Q6 | GGUF | — | — | — | — | — | HumanEval 157/164 (95.73%); HumanEval+ 150/164 (91.46%) | reference not stated | L106 @ main | |
| unsloth/GGUF measured by wiklif29.1 GB only in table header, no candidate [L65] | Q8_0 | GGUF | — | — | — | — | — | GPQA Diamond 92.4 ± 3.7; IFBench_loose 81.7; IFBench_strict 79.0; LiveCodeBench v6 88.5 ± 5.5 | vs a Q8_0 | L67 @ main | |
| unsloth/NVFP4 measured by shisa-aiMeasured locally at revision 7d6f8d4d; median KLD c21 [L58] | unsloth/Qwen3.8-27B-NVFP4 | NVFP4 | 23.42 | download | 0.030155 | 93.14% | — | — | kld_99 0.335824 | vs BF16 | L58 @ main |
| unsloth/NVFP4 measured by QUASAR-QAT168/496 linears NVFP4 [L84] | Unsloth NVFP4 | NVFP4 | 23.40 | download | — | — | — | — | GPQA-Diamond 89.39; AIME'26 97.78 | vs BF16 | L84 @ main |
| unsloth/NVFP4 measured by TelperionAI[L59] | unsloth/Qwen3.8-27B-NVFP4 | NVFP4 | 23.40 | download | — | 91.75% | — | — | divmed 19 | vs BF16 | L59 @ main |
| unsloth/NVFP4 measured by TelperionAI[L60] | unsloth/Qwen3.8-27B-NVFP4 | NVFP4 | 23.40 | download | — | 91.75% | — | — | — | vs BF16 | L60 @ main |
| unsloth/NVFP4 measured by TelperionAI | unsloth/Qwen3.8-27B-NVFP4 | NVFP4 | 23.40 | download | — | 91.75% | — | — | — | vs BF16 | L59 @ main |
| unsloth/NVFP4 measured by malaiwahshard-0 of v5 suite [L503-508]resident weights, not serialized [L487] | unsloth/Qwen3.8-27B-NVFP4 | NVFP4 | 22.91 | resident | 0.030115 | 93.16 % | — | — | kld:v3 0.094978; top1:v3 90.53 % | vs BF16 | L396 @ main |
| unsloth/NVFP4 measured by malaiwah[L782] | unsloth/Qwen3.8-27B-NVFP4 v3 136-context | NVFP4 | 22.91 | resident | 0.094978 | 90.53 % | — | — | — | other or no reference | L782 @ main |
| unsloth/NVFP4 quoted by acyildirimer, not measured[L81] | unsloth/Qwen3.8-27B-NVFP4 | NVFP4 | — | — | — | — | — | MMLU 84.3; ARC-C 96.9; HellaSwag 94.3; GSM8K 97.1; HumanEval 89.6; 4-task mean 93.2 | vs BF16 | L81 @ main | |
| unsloth/NVFP4 measured by malaiwahresident 21.34 GiB measured, not serialized [L486]; shard-0 [L463] | unsloth/Qwen3.8-27B-NVFP4 | NVFP4 | — | 0.030115 | 93.16 % | — | — | — | vs BF16 | L463 @ main | |
| unsloth/NVFP4 measured by malaiwah[L503]; resident 21.34 GiB in prose [L522] | unsloth/Qwen3.8-27B-NVFP4 shard0 | NVFP4 | — | 0.030115 | 93.16 % | — | — | — | other or no reference | L503 @ main | |
| unsloth (link truncated in prompt) measured by primitive-ai[L84] | unsloth NVFP4 | NVFP4 | 23.41 | download | — | — | — | — | overall 89.6; knowledge 91.4; call 85.0; abstain 57.5 | vs BF16 | L84 @ main |
| unsloth (repo not named on card) measured by talxccbaseline row; General Benches cell is 'Baseline (100%)' [L63] | Unsloth Q4_K_M (Standard) | 15.20 | download | — | — | — | — | Coding_Retention 100% (Baseline) | other or no reference | L63 @ main | |
| unsloth (repo not named on card) measured by AtomicChat[L205] | UD-Q4_K_XL | — | 0.00955 | — | — | — | — | vs BF16 | L205 @ main | ||
| varshavvvvv/Ternary-CyberX-2-27B-gguf reference [L212]publisher comparison build | Qwen3.8-27B FP16 | GGUF | 54.00 | text weights (excl. vision, MTP) | — | — | — | — | MMLU-Redux 91.46; MuSR 79.63; GSM8K 97.19; MATH-500 99.80; AIME25 96.67; AIME26 94.58; HumanEval+ 93.29; MBPP+ 83.86; LiveCodeBench 90.05; IFEval 91.50; BFCL v3 76.74; MMMU-Pro 81.73; OCR Bench v2 60.99 | other or no reference | L212 @ main |
| Vontra/MLX-4bit measured by anubhav200[L81] | Vontra/Qwen3.8-27B-MLX-4bit (MLX affine 4-bit) | MLX | — | — | 92.80% | — | — | — | vs BF16 | L81 @ main | |
| WarlaxZ/NInfer-3090 Publisher's variant [L89]publisher comparison build | This layout with a 6-bit embedding | other | — | 0.0304 | — | — | — | — | other or no reference | L89 @ main | |
| WaveCut/MLX-4bit-DWQ baseline mlx-community 4-bit RTN [L26,L50]publisher comparison build | Original 4-bit RTN baseline | MLX | — | 0.204094 | — | — | — | — | other or no reference | L50 @ main | |
| XpressAI/RYS-GGUF base model quant not stated in table [L41-47]publisher comparison build | Base | GGUF | — | — | — | — | — | Reasoning probe 58.82%; EQ-140 82.65; BFCL micro 82.60%; BFCL macro 80.54%; Math probe 1.000 | other or no reference | L43 @ main | |
| YourHighnessLA/L0xRE-27b-Low-MTP [L72]publisher comparison build | native-quant baseline | other | — | — | — | — | — | internal_suite_pass@1 117 / 150 | other or no reference | L72 @ main | |
| Youssofal/MTPLX-Bare-Speed 2,389-position teacher-forced set [L60]relation of Bare Speed to Optimized Speed not statedpublisher comparison build | Optimized Speed (4-bit dynamic) | other | — | 0.012 | — | — | — | — | vs BF16 | L60 @ main | |
| Youssofal/MTPLX-Bare-Speed [L60]publisher comparison build | Optimized Quality (8-bit dynamic) | other | — | — | 99.3 percent | — | — | — | vs BF16 | L60 @ main | |
| Youssofal/MTPLX-Bare-Speed-FP16 sibling pack, not this repo [L69,L106]publisher comparison build | Optimized Speed (4-bit dynamic) | other | — | 0.012 | — | — | — | — | vs BF16 | L69 @ main | |
| Youssofal/MTPLX-Bare-Speed-FP16 sibling pack, not this repo [L69,L106]publisher comparison build | Optimized Quality (8-bit dynamic) | other | — | KL 0.0005. Exactness on this release: a | — | — | — | — | vs BF16 | L69 @ main | |
| Youssofal/MTPLX-Optimized-Quality Sibling pack by same publisher; KL type (mean/median) not stated [L59]publisher comparison build | Optimized Speed (4-bit dynamic) | other | — | — | — | — | — | kld:code_prose_json_multilingual 0.012; top1:code_prose_json_multilingual 96.0 percent | vs BF16 | L59 @ main | |
| Youssofal/MTPLX-Optimized-Quality-FP16 99.3 percent top-1 agreement in prose [L68]publisher comparison build | Optimized Quality (8-bit dynamic) parent | other | — | KL 0.0005. Exactness on this release: a | 99.3 percent | — | — | — | vs BF16 | L68 @ main | |
| Youssofal/MTPLX-Optimized-Quality-FP16 96.0 percent top-1 agreement in prose [L68]publisher comparison build | Optimized Speed (4-bit dynamic) | other | — | 0.012 | — | — | — | — | vs BF16 | L68 @ main | |
| Youssofal/MTPLX-Optimized-Speed separate repo Youssofal/Qwen3.8-27B-MTPLX-Optimized-Quality [L130]publisher comparison build | Optimized Quality (8-bit dynamic) | other | — | KL 0.0005. Exactness on this release: a | 99.3 percent | — | — | — | vs BF16 | L58 @ main | |
| Youssofal/MTPLX-Optimized-Speed-FP16 parent pack figures [L68]publisher comparison build | Optimized Quality (8-bit dynamic) | other | — | — | 99.3 percent | — | — | — | vs BF16 | L68 @ main | |
| Zynerji/Ektome-PristinelyUncensored sibling repo; max context on 24 GB is a context length not a quality value [L44]quality figures only in prose [L54]publisher comparison build | HOMEUSER-16-24 | other | 14.70 | download | — | — | — | — | MMLU 0.7995 | other or no reference | L44 @ main |
| Zynerji/Ektome-PristinelyUncensored-GPTQ-MTP [L147-148]publisher comparison build | bf16 parent | GPTQ | — | — | — | — | — | MMLU-val 0.8175; RefusalCompliance 1.000 | vs BF16 | L147 @ main | |
| Zynerji/Ektome-PristinelyUncensored-GPTQ-MTP bf16 parent receipts [L75-76]publisher comparison build | edited (bf16, post-edit) | GPTQ | — | — | — | — | — | RefusalCompliance 1.000 | vs BF16 | L75 @ main |
Publisher notes
Method, evaluation setup and problems we found, per repository. Line numbers refer to the card snapshot linked at the bottom of each note.
0xBakeer/TandemLLM-Qwen3.8-27B-EXL3-4.5
- Reference
- BF16 Qwen3.8-27B
- Eval data
- wikitext-2
- Context
- not stated
- Tool
- exllamav3 1.6.0
Card snapshot main · updated 2026-10-13
0xBakeer/TandemLLM-Qwen3.8-27B-EXL3-6.0
- Reference
- BF16 model (Qwen/Qwen3.8-27B)
- Eval data
- wikitext-2
- Context
- not stated
- Tool
- exllamav3 1.6.0 (raw output in quality/)
Card snapshot main · updated 2026-10-13
0xBakeer/TandemLLM-Qwen3.8-27B-NVFP4
- Reference
- Qwen/Qwen3.8-27B BF16, change in loss per token, teacher-forced
- Eval data
- four held-out texts (prose 1,521 tok; Triton Python 8,002 tok; Python 10,500 tok; Wikipedia 11,999 tok); gate reads first 2,048 tokens
- Context
- 2,048 for gate
- Tool
- GPTQ on NVFP4 grid, TandemLLM tools/quality_gate.py
Card snapshot main · updated 2026-09-29
0xWhiteMage/Qwen3.8-27B-Kearuga
- Reference
- native BF16 Qwen/Qwen3.8-27B
- Eval data
- Fidelity-40 (40 prompts); held-out set of 72,082 tokens; Quality-200 (GSM8K, HumanEval, IFEval, agentic coding)
- Context
- not stated
- Tool
- SGLang on DGX Spark, KV cache BF16, thinking off, temperature 0
Card snapshot main · updated 2026-09-24
1bit-MONSTER/Qwen3.8-27B-Q4_0-H32-GGUF
- Reference
- BF16
- Eval data
- wikitext-2 (40 x 512)
- Context
- 512
- Tool
- llama-bench for pp512; KLD via 1bit engine on Strix Halo
Card snapshot main · updated 2026-09-28
1bit-MONSTER/Qwen3.8-27B-UD-Q4_K_XL-MTP
- Reference
- the reference (not specified further)
- Eval data
- not stated
- Context
- not stated
- Tool
- not stated (measured on Strix Halo, Vulkan)
Card snapshot main · updated 2026-09-26
6block/Qwen3.8-27B-GGUF
- Reference
- master BF16 (not in this repo)
- Eval data
- not stated
- Context
- not stated
- Tool
- not stated (in-house fixed setup, same command for all tiers)
Card snapshot main · updated 2026-08-19
abhishekchohan/Qwen3.8-27B-AWQ-INT4-FP8KV
- Reference
- abhishekchohan/Qwen3.8-27B-AWQ-INT4 with bf16 KV cache (same weights, only KV dtype differs)
- Eval data
- GSM8K (1319), GPQA Diamond (198), Minerva Math500 (500), ARC-Easy (2376), ARC-Challenge (1172), HellaSwag (2000), Winogrande (1267), TruthfulQA MC1 (817); needle-in-a-haystack
- Context
- fp8-e4m3 KV; KV scales calibrated on 112 sequences x 262,144 tokens; needle test up to 131,072
- Tool
- same harness as the AWQ-INT4 card; thinking generative and loglikelihood protocols
Card snapshot main · updated 2026-08-27
abhishekchohan/Qwen3.8-27B-GPTQ-INT4
- Reference
- BF16 base
- Eval data
- GSM8K (250-sample subset), GPQA Diamond (198), RealWorldQA (765), MathVision (3040), ARC-Easy (2376)
- Context
- not stated for evals; needle test up to 262,144
- Tool
- lm-eval-harness, hf-multimodal backend, greedy; thinking enabled for generative tasks, loglikelihood for ARC-Easy
Card snapshot main · updated 2026-08-25
abhishekchohan/Qwen3.8-27B-GPTQ-INT4-FP8KV
- Reference
- Base BF16 (bf16 KV)
- Eval data
- lm-evaluation-harness tasks (GSM8K, GPQA Diamond, Minerva Math500, RealWorldQA, MathVision, ARC, HellaSwag, Winogrande, TruthfulQA MC1)
- Context
- max_gen_toks=8192; thinking enabled; greedy, seed 1234
- Tool
- lm-evaluation-harness with vLLM 0.27.1 on RTX PRO 6000 Blackwell
Card snapshot main · updated 2026-10-09
abihsoro/Qwen3.8-27B-AWQ-INT4
- Reference
- no BF16 reference; compared against cyankiwi/Qwen3.6-27B-AWQ-INT4 (different base model)
- Eval data
- GSM8K, 100 samples; calibration HuggingFaceH4/ultrachat_200k 256 samples
- Context
- max_gen_toks=2048
- Tool
- lm-evaluation-harness (local-chat-completions); llm-compressor
Card snapshot main · updated 2026-08-15
abuarchive/Qwen3.8-27B-2bit-oQ2
- Reference
- official 4-bit release (mlx-community/Qwen3.8-27B-4bit), same hardware and settings
- Eval data
- allenai/tulu-3-sft-mixture, 40 held-out samples, seed 42; plus 16-prompt battery
- Context
- sequence length 256
- Tool
- MLX perplexity; sensitivity-pass per-layer bit allocation
Card snapshot main · updated 2026-08-16
acyildirimer/Qwen3.8-27B-AutoQuant-NVFP4-FP8-5bit
- Reference
- Qwen/Qwen3.8-27B BF16 (comparison row)
- Eval data
- MMLU 5-shot (14,042 q), ARC-C, HellaSwag 10-shot (10,042 q), GSM8K, HumanEval
- Context
- not stated
- Tool
- vLLM chat completions, thinking disabled; comparison rows from public witcheer/rtx-5090-benchmarks, serving stacks differ
Card snapshot main · updated 2026-08-30
airagrp/Qwen3.8-27B-mlx-mxfp8-L
- Reference
- bf16 Qwen3.8-27B
- Eval data
- WikiText-2 raw test split, 297,053 tokens
- Context
- 2048-token windows, fresh KV cache per window
- Tool
- benchmark_ppl_kld.py (mlx-vlm 0.6.17)
Card snapshot main · updated 2026-09-04
airagrp/Qwen3.8-27B-mlx-mxfp8-M
- Reference
- bf16 Qwen3.8-27B
- Eval data
- wikitext-2-raw-v1 test split, 297,053 tokens
- Context
- 2048-token windows, fresh KV cache per window
- Tool
- benchmark_ppl_kld.py, mlx-vlm 0.6.17; full-vocabulary mean D_KL(p_bf16 || p_model) in nats/token
Card snapshot main · updated 2026-09-04
airagrp/Qwen3.8-27B-mlx-mxfp8-S
- Reference
- bf16 Qwen3.8-27B
- Eval data
- WikiText-2 raw test split (297,053 tokens)
- Context
- 2048-token windows, fresh KV cache per window
- Tool
- benchmark_ppl_kld.py, mlx-vlm 0.6.17
Card snapshot main · updated 2026-09-04
airagrp/Qwen3.8-27B-mlx-mxfp8-XL
- Reference
- bf16 Qwen3.8-27B
- Eval data
- WikiText-2 raw test split (297,053 tokens)
- Context
- 2048-token windows
- Tool
- benchmark_ppl_kld.py with mlx-vlm 0.6.17
Card snapshot main · updated 2026-09-04
airagrp/Qwen3.8-27B-mlx-nvfp4-L
- Reference
- bf16 Qwen3.8-27B, mean per-token D_KL(p_bf16 || p_model) over full vocabulary
- Eval data
- WikiText-2 raw test split, 297,053 tokens
- Context
- 2048-token windows, fresh KV cache per window
- Tool
- benchmark_ppl_kld.py, mlx-vlm 0.6.17
Card snapshot main · updated 2026-09-04
airagrp/Qwen3.8-27B-mlx-nvfp4-M
- Reference
- bf16 Qwen3.8-27B
- Eval data
- WikiText-2 raw test split, 297,053 tokens
- Context
- 2048-token windows, fresh KV cache per window
- Tool
- benchmark_ppl_kld.py, mlx-vlm 0.6.17
Card snapshot main · updated 2026-09-04
airagrp/Qwen3.8-27B-mlx-nvfp4-S
- Reference
- bf16 Qwen3.8-27B
- Eval data
- WikiText-2 raw test split, 297,053 tokens
- Context
- 2048-token windows, fresh KV cache per window
- Tool
- benchmark_ppl_kld.py, mlx-vlm 0.6.17
Card snapshot main · updated 2026-09-04
airagrp/Qwen3.8-27B-mlx-nvfp4-XL
- Reference
- bf16 Qwen3.8-27B
- Eval data
- WikiText-2 raw test split, 297,053 tokens
- Context
- 2048-token windows, fresh KV cache per window
- Tool
- benchmark_ppl_kld.py, mlx-vlm 0.6.17, benchmarked 2026-09-03
Card snapshot main · updated 2026-09-04
aj9o9/Qwen3.8-27B-GGUF
- Reference
- official BF16 logits
- Eval data
- wiki.test.raw (KLD); author's locked GLSL / hard suite, think off, t=0
- Context
- n_ctx=512
- Tool
- llama.cpp
Card snapshot main · updated 2026-08-17
akopytko/Qwen3.8-27B-NVFP4-GGUF
- Reference
- base bf16 model
- Eval data
- wikitext test set
- Context
- not stated
- Tool
- KLD and Top P agreement (llama.cpp style); base perplexity 6.7945
Card snapshot main · updated 2026-10-04
Alogotron/Qwen3.8-27B-MixQ4A6E8-GGUF
- Reference
- BF16 base logits (freshly generated)
- Eval data
- WikiText-2 test, full set, 602 chunks
- Context
- 512 (ctx/batch/ubatch 512)
- Tool
- llama.cpp b10442 Vulkan perplexity lane
Card snapshot main · updated 2026-08-19
Alogotron/Qwen3.8-27B-MixQ5A6E8-GGUF
- Reference
- BF16 (BF16-relative KL; BF16 PPL 6.648031)
- Eval data
- full WikiText-2 PPL
- Context
- not stated
- Tool
- llama.cpp b10442 Vulkan perplexity lane; Bartowski imatrix used for quantization
Card snapshot main · updated 2026-08-19
amd/Qwen3.8-27B-Quark-AWQ-INT4-W4A16
- Reference
- BF16 Qwen/Qwen3.8-27B (measured by the card author)
- Eval data
- GSM8K 5-shot; Wikitext perplexity; BFCL single_turn
- Context
- not stated
- Tool
- lm-evaluation-harness (vLLM backend); bfcl_eval; Quark AWQ with 128 calibration samples, seq len 512
Card snapshot main · updated 2026-08-28
amd/Qwen3.8-27B-Quark-Qronos-INT4-W4A16
- Reference
- Qwen/Qwen3.8-27B BF16, measured by the card author
- Eval data
- GSM8K 5-shot (thinking and no-thinking), Wikitext perplexity, BFCL single_turn
- Context
- not stated
- Tool
- lm-evaluation-harness (vLLM backend), bfcl_eval; Quark Qronos INT4 W4A16 group 128, 128 calibration samples x 512 tokens
Card snapshot main · updated 2026-08-28
andrew-paul/Qwen3.8-27B-GGUF
- Reference
- BF16 (28Q score = 100%)
- Eval data
- not stated for KLD; 28Q screening metric
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-09-29
anemll/anemll-forge-qwen3.8-27B
- Reference
- BF16 Qwen3.8-27B (teacher-forced KL, top-512 plus tail on ANE trace)
- Eval data
- KL-512: 64 chats (40,023 positions); held-out chats 37,076; WikiText 16x1,024; last 2K of 16 x 16,384-token windows from 12 long sessions
- Context
- short chat traces (longest 857 tokens); 8K/64K ppl; 16K windows
- Tool
- ANEMLL harness on M6 ANE and PyTorch; GGUFs scored with llama.cpp on M3 Ultra
Card snapshot main · updated 2026-10-10
anemll/anemll-quantized-qwen3.8-27b-for-CoreAI
- Reference
- BF16 Qwen3.8-27B (teacher-forced KL)
- Eval data
- 64 chats (40,023 positions), held-out chats (37,076), WikiText 16x1,024, last 2K of 16 x 16,384-token windows from 12 long sessions
- Context
- short chat traces and 16K windows
- Tool
- PyTorch for ANEMLL and Mirai S; llama.cpp on M3 Ultra for GGUFs
Card snapshot main · updated 2026-10-10
anubhav200/Qwen3.8-27B-exl3-4.00bpw-bf16vision
- Reference
- 6.00 bpw EXL3 model (near-full-precision, itself EXL3)
- Eval data
- 61,575 thinking-mode response tokens generated by the 6.00 bpw EXL3 model (cal_trace.json)
- Context
- not stated
- Tool
- TensorFold forward (fp8 KV, bf16 prompt activations), DGX Spark
Card snapshot main · updated 2026-10-11
Ar3ntiz/qwen3_27B_side8
- Reference
- original Qwen3.8-27B (BF16)
- Eval data
- 16 x 512 held-out positions (KL/top-1); WikiText-2 test (PPL); 120 graded questions (20 per subset) for answer-level eval
- Context
- 512-token positions
- Tool
- SUC runtime (KL/top-1 vs original); greedy answer eval
Card snapshot main · updated 2026-10-17
armin1/Qwen3.8-27B-NVFP4-FP8-Mixed-LH
- Reference
- BF16 original (teacher-forced)
- Eval data
- 58 documents of code, prose, long context (16K/32K) and chat, 110,155 positions; GSM8K test (1,319 questions) for reasoning
- Context
- documents up to 16K/32K; GSM8K 16K cap
- Tool
- SGLang on DGX Spark; KL over top-20 tokens (lower bound on full KL)
Card snapshot main · updated 2026-10-02
armin1/Qwen3.8-27B-NVFP4-FP8-Mixed-LH-Heretic
- Reference
- BF16 original, teacher-forced
- Eval data
- 58 documents (code, prose, long context 16K/32K, chat), 110,155 positions; GSM8K test 1,319 questions thinking on; quantization calibration Nemotron post-training v3 (512 x 2,048)
- Context
- GSM8K 16K cap; KL computed over top-20 tokens (lower bound on full KL)
- Tool
- NVIDIA ModelOpt hf_ptq.py local_hessian; SGLang on DGX Spark; Heretic for abliteration
Card snapshot main · updated 2026-10-02
atakans/Qwen3.8-27B-NVFP4-BF16head
- Reference
- BF16 (perplexity 8.30)
- Eval data
- WikiText-2 test, 100 x 512 tokens, top-20 logprobs
- Context
- 512
- Tool
- ModelOpt 0.47.0 Local-Hessian calibration; vLLM 0.30.0 measurement script
Card snapshot main · updated 2026-10-14
AtomicChat/Qwen3.8-27B-GGUF
- Reference
- original BF16 weights converted to GGUF, run unquantized (PPL 4.5219 +/- 0.0238)
- Eval data
- eval_neutral held-out split of AtomicChat/calib-corpora, 87 chunks
- Context
- 4096
- Tool
- llama-perplexity --kl-divergence
Card snapshot main · updated 2026-08-18
AveLabs/Qwen3.8-27B-Perplexing-Pegasus-NVFP4
- Reference
- Decisions: original BF16 pplx-decider; chat: stock Qwen3.8-27B in its official NVFP4 build
- Eval data
- Decisions: JevBench public, WANLI, Open-Jev v1.1 mix; chat: IFEval, GSM8K zero-shot CoT, MMLU-Pro 1,008 questions; calibration 1,024 x 512 tokens of cnn_dailymail
- Context
- MMLU-Pro 16K-token answer budget, unanswered reruns at 64K
- Tool
- SGLang, lm-evaluation-harness 0.4.13, reasoning effort low, temperature 0; quantized with NVIDIA ModelOpt
Card snapshot main · updated 2026-10-05
Avesed/Qwen3.8-27B-INT4-W4A16
- Reference
- none (absolute accuracy, no bf16 row)
- Eval data
- GSM8K, 250 questions
- Context
- max_tokens=24576
- Tool
- vLLM tp2 over OpenAI API, thinking on, temp 1.0 top_p 0.95 top_k 20, boxed-answer extraction
Card snapshot main · updated 2026-08-16
Avesed/Qwen3.8-27B-INT8-W8A8
- Reference
- none (absolute accuracy, no bf16 row)
- Eval data
- GSM8K, 250 questions
- Context
- max_tokens=24576
- Tool
- vLLM tp2 over OpenAI API, thinking on, temp 1.0 top_p 0.95 top_k 20, boxed-answer extraction
Card snapshot main · updated 2026-08-16
avlp12/Qwen3.8-27B-Alis-MLX-4bit
- Reference
- bf16 (Qwen/Qwen3.8-27B); KL is exact full-vocab (248,320-way fp32 softmax) to bf16
- Eval data
- PPL: wikitext-2 test (en, 35,981 tok), Korean Wikipedia (ko, 35,179 tok), CPython stdlib (code, 32,295 tok); KL: ~100K en+ko+code tokens
- Context
- PPL window 2048 stride 512 (~103K scored tokens); KL paired non-overlapping ctx-2048 windows
- Tool
- MLX strided PPL harness with paired blockwise SE (512-token blocks); exact-KL sweep
Card snapshot main · updated 2026-08-22
avlp12/Qwen3.8-27B-Alis-MLX-6bit
- Reference
- bf16 original Qwen/Qwen3.8-27B
- Eval data
- wikitext-2 test (en, 35,981 tok), Korean Wikipedia (ko, 35,179 tok), CPython stdlib (code, 32,295 tok); exact-KL sweep over ~100K en+ko+code tokens
- Context
- ctx 2048, stride 512 (PPL); ctx-2048 windows (KL); ctx 1024 x35 windows wikitext-2 (cross-check)
- Tool
- MLX strided PPL with paired per-token excess NLL, blockwise SE; full-vocab fp32 KL
Card snapshot main · updated 2026-08-22
avlp12/Qwen3.8-27B-Alis-MLX-8bit
- Reference
- bf16 original (Qwen/Qwen3.8-27B)
- Eval data
- wikitext-2 test (en, 35,981 tok), Korean Wikipedia (35,179 tok), CPython stdlib (code, 32,295 tok); KL sweep over ~100K en+ko+code tokens
- Context
- window 2048, stride 512 (PPL); ctx 2048 non-overlapping windows (KL)
- Tool
- MLX strided perplexity, paired blockwise SE; exact full-vocab KL
Card snapshot main · updated 2026-08-22
avyukth/Qwen3.8-27B-AWQ-INT4
- Reference
- not stated
- Eval data
- GSM8K, 40 problems only, exact match, temperature 0
- Context
- max_model_len=8192
- Tool
- vLLM on RTX 3090
Card snapshot main · updated 2026-08-15
AxionML/Qwen3.8-27B-NVFP4
- Reference
- not stated
- Eval data
- calibration abisee/cnn_dailymail train, 1,024 samples; eval GSM8K full 1,319 split and Terminal-Bench 2.1 84-task subset
- Context
- calibration seq length 512
- Tool
- NVIDIA Model Optimizer commit 87c9f8cf; SGLang TP4 for evals
Card snapshot main · updated 2026-09-29
ayushcluster00/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF-Config
- Reference
- Qwen3.8-27B base (KL vs base of the abliterated weights, not a quant comparison)
- Eval data
- 100 harmful-behaviors prompts (refusals); imatrix over wikitext-2-raw / tiny_shakespeare
- Context
- not stated
- Tool
- heretic ARA
Card snapshot main · updated 2026-08-18
baa-ai/Qwen3.8-27B-RAM-13GB-GGUF
- Reference
- not stated (card claims no loss vs full-precision model)
- Eval data
- WikiText-2 (perplexity); GSM8K 50 questions; MMLU-math 60 questions
- Context
- not stated
- Tool
- llama.cpp (llama-server for context tests)
Card snapshot main · updated 2026-10-16
baa-ai/Qwen3.8-27B-RAM-24GB-MLX
- Reference
- Qwen3.8-27B BF16 (full precision)
- Eval data
- MMLU 100-question calibrated subset spanning 57 subjects
- Context
- not stated
- Tool
- identical harness for both models, enable_thinking=False; agent_safety_gate.py and canary.py for safety screen
Card snapshot main · updated 2026-09-27
baa-ai/Qwen3.8-27B-RAM-31GB-GGUF
- Reference
- BF16 (for the sibling MLX MMLU claim)
- Eval data
- imatrix: 100 chunks wikitext-2 + 200 MMLU-Pro questions
- Context
- not stated
- Tool
- llama-quantize with imatrix
Card snapshot main · updated 2026-08-20
bartowski/Qwen3.8-27B-GGUF
- Reference
- bf16
- Eval data
- wiki.test.raw
- Context
- not stated (100 chunks)
- Tool
- llama.cpp b10896 (perplexity / KL divergence)
Card snapshot main · updated 2026-10-13
bernhardbrieger/Qwen3.8-27B-GPTQ-Int4
- Reference
- none; absolute perplexity only, no base-vs-quant comparison
- Eval data
- single 260-token sample
- Context
- not stated
- Tool
- scripts/measure_ppl_quant.py on vLLM server
Card snapshot main · updated 2026-10-06
bernhardbrieger/Qwen3.8-27B-NVFP4
- Reference
- none (absolute PPL only; no base-vs-quant delta reported)
- Eval data
- single 260-token sample, not otherwise named
- Context
- not stated
- Tool
- scripts/measure_ppl_quant.py
Card snapshot main · updated 2026-09-23
bg-digitalservices/Qwen3.8-27B-ROCmFP4-STRIX-GGUF
- Reference
- own F16 conversion of the same weights
- Eval data
- wikitext-2 (EN) and held-out German Wikipedia slice (DE), disjoint from calibration corpus; 200 chunks
- Context
- not stated
- Tool
- KL-divergence (llama.cpp perplexity-style)
Card snapshot main · updated 2026-08-19
billmccartney/Qwen3.8-27B-exl3-8.00bpw
- Reference
- cached BF16 distributions (Qwen3.8-27B BF16)
- Eval data
- WikiText2 test: 100 rows x 2,048 tokens (PPL r-100); KL over 204,800 logit positions
- Context
- 2,048-token rows, stride 2,048, no chat wrapper
- Tool
- exllamav3 1.4.2 qbench
Card snapshot main · updated 2026-08-17
bowmanslayer/Qwen3.8-27B-W4A16-vision
- Reference
- none (no BF16 comparison column)
- Eval data
- MMLU, CMMLU, C-Eval, ARC-C, TruthfulQA MC1, GSM8K, MATH-500, BBH, HumanEval, IFEval strict; samples of 100-164 each
- Context
- not stated
- Tool
- thinking ON, sampled subsets; truncation-corrected accuracy also reported
Card snapshot main · updated 2026-08-17
btbtyler09/Qwen3.8-27B-GPTQ-4bit
- Reference
- BF16 original
- Eval data
- wikitext-2-raw-v1 test
- Context
- seq_len=2048, stride=512
- Tool
- not stated (GPTQModel v6.0.3 for quantization)
Card snapshot main · updated 2026-08-14
btbtyler09/Qwen3.8-27B-GPTQ-8bit
- Reference
- BF16 original
- Eval data
- wikitext-2-raw-v1 test set; calibration evol-codealpaca-v1 + C4, 256 samples
- Context
- seq_len 2048, stride 512
- Tool
- GPTQ 8-bit group size 32
Card snapshot main · updated 2026-08-14
Bucoid/Qwen3.8-27B-Heretic-Ara-16GB-VRAM-IQ4-XS-MTP-GGUF
- Reference
- Heretic-Ara BF16 (base)
- Eval data
- not stated
- Context
- not stated
- Tool
- not stated (PPL/KLD table, llama.cpp-style output)
Card snapshot main · updated 2026-08-23
Bucoid/Qwen3.8-27B-IQ4-XS-MTP-16GB-VRAM-GGUF
- Reference
- base model (unquantized)
- Eval data
- not stated
- Context
- not stated
- Tool
- not stated (perplexity/KLD tooling)
Card snapshot main · updated 2026-08-16
Bucoid/Qwen3.8-27B-Uncensored-IQ4-XS-MTP-16GB-VRAM-GGUF
- Reference
- base Qwen3.8-27B Uncensored (precision and file not stated)
- Eval data
- not stated
- Context
- not stated
- Tool
- llama.cpp-style perplexity/KLD output (Mean PPL, Mean KLD, Same top p, Delta p); version not stated
Card snapshot main · updated 2026-08-18
cafonez/Qwen3.8-27B-ROCmI4-MTP-GGUF
- Reference
- local Q4_K_M reference
- Eval data
- not stated
- Context
- not stated
- Tool
- ROCmFPX build of llama.cpp on gfx1151
Card snapshot main · updated 2026-08-22
caiovicentino1/Qwen3.8-27B-GPTQ-gs64-FOEM
- Reference
- none (no BF16 baseline run)
- Eval data
- C4 perplexity 40 docs / 16,683 tokens; HellaSwag 20, ARC-C 20, MMLU 8 q x 4 subjects, GSM8K 12 samples
- Context
- not stated
- Tool
- gptqmodel 6.0.3 + transformers 5.15.0, RTX PRO 6000
Card snapshot main · updated 2026-08-15
Capicua25x/Qwen3.8-27B-MXFP4-Quark-RDNA4
- Reference
- same checkpoint served in bf16 by a cloud provider (ref)
- Eval data
- AA class-A paired items: GSM8K, IFEval, AA-LCR, GPQA-Diamond, AIME'25, tau2, MMLU-Pro, Terminal-Bench hard, HLE (n in row labels)
- Context
- not stated
- Tool
- AMD Quark MXFP4, seed 1234, on-spec sampling, thinking ON except nothink rows
Card snapshot main · updated 2026-08-25
cdiamond/Qwen3.8-27B-iMatrix-NVFP4-MTP-GGUF
- Reference
- none (perplexity only)
- Eval data
- short WikiText-2 control
- Context
- not stated
- Tool
- llama.cpp (llama-imatrix for calibration)
Card snapshot main · updated 2026-08-17
Chungulus/Qwen3.8-27B-EXL3
- Reference
- not stated (full-vocabulary held-out comparison, reference presumably unquantized)
- Eval data
- held-out comparison, 512 token positions
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-08-16
Chungulus/Qwen3.8-27B-FP8
- Reference
- not stated (full-vocabulary held-out comparison, presumably vs source model)
- Eval data
- held-out comparison on 64 token positions; calibration corpus CC0 authored for project, 216 text/tool + 8 vision samples (None used for FP8 build)
- Context
- not stated
- Tool
- GPTQModel FP8 (float8_e4m3fn, RTN fallback config), RTX 4070 12 GB
Card snapshot main · updated 2026-08-18
Chungulus/Qwen3.8-27B-GPTQ-INT4-G128
- Reference
- not stated
- Eval data
- full-vocabulary held-out comparison on 512 token positions; calibration corpus 128 samples (core.jsonl:128), original CC0 material
- Context
- calibration sequence length 2048
- Tool
- GPTQModel 7.3.4+
Card snapshot main · updated 2026-08-16
Chungulus/Qwen3.8-27B-GPTQ-INT4-G32
- Reference
- not stated (held-out comparison of full-vocabulary logits; presumably unquantized)
- Eval data
- not stated (held-out)
- Context
- 512 token positions
- Tool
- full-vocabulary held-out comparison; GPTQModel quantizer
Card snapshot main · updated 2026-08-16
Chungulus/Qwen3.8-27B-GPTQ-INT8
- Reference
- not stated (full-vocabulary comparison, presumably against the base model)
- Eval data
- held-out comparison; calibration corpus original CC0 material, 128 samples (calibration/core.jsonl:128)
- Context
- 512 token positions; calibration seqlen 2048
- Tool
- GPTQModel 7.3.4+, GPTQ 8-bit group size 128 desc_act
Card snapshot main · updated 2026-08-16
Chungulus/Qwen3.8-27B-MLX-3bit-Group32
- Reference
- BF16 (fixed-logit comparison)
- Eval data
- fixed original text, 106 positions, not a public benchmark
- Context
- not stated (longest validation prompt 73 tokens)
- Tool
- mlx-vlm 0.6.1 validation_result.json
Card snapshot main · updated 2026-08-16
Chungulus/Qwen3.8-27B-MLX-4bit
- Reference
- not stated (semantic similarity proxy against reference outputs)
- Eval data
- short validation prompts, longest 73 prompt tokens
- Context
- 73 prompt tokens longest
- Tool
- mlx-vlm 0.6.1, validation_result.json; calibration none
Card snapshot main · updated 2026-08-15
Chungulus/Qwen3.8-27B-MLX-4bit-Group32
- Reference
- BF16 fixed logits
- Eval data
- fixed original text, 106 positions (not a public benchmark)
- Context
- longest validation prompt 73 tokens
- Tool
- mlx-vlm 0.6.1
Card snapshot main · updated 2026-08-16
Chungulus/Qwen3.8-27B-MLX-5bit-Group128
- Reference
- BF16 fixed-logit comparison, float16-stored log-probs after float32 log-softmax
- Eval data
- fixed original text, 106 positions (not a public benchmark)
- Context
- not stated (longest validation prompt 73 tokens)
- Tool
- mlx-vlm 0.6.1, MLX affine 5-bit group 128, no calibration
Card snapshot main · updated 2026-08-16
Chungulus/Qwen3.8-27B-MLX-5bit-Group32
- Reference
- BF16 fixed-logit comparison
- Eval data
- fixed original text, 106 positions; not a public benchmark
- Context
- longest validation prompt 73 prompt tokens
- Tool
- mlx-vlm 0.6.1; MLX affine 5-bit group size 32, no calibration; BF16 log-probs stored in float16 after float32 log-softmax
Card snapshot main · updated 2026-08-16
Chungulus/Qwen3.8-27B-MLX-6bit-Group32
- Reference
- BF16 fixed logits
- Eval data
- fixed original text, 106 positions, not a public benchmark
- Context
- longest validation prompt 73 tokens
- Tool
- mlx-vlm 0.6.1, MLX affine 6-bit group 32
Card snapshot main · updated 2026-08-16
Chungulus/Qwen3.8-27B-MLX-8bit-Group32
- Reference
- BF16 fixed-logit comparison on fixed original text, 106 positions; BF16 log-probs stored as float16
- Eval data
- fixed original text, not a public benchmark
- Context
- not stated
- Tool
- mlx-vlm 0.6.1 validation_result.json
Card snapshot main · updated 2026-08-16
Chungulus/Qwen3.8-27B-MLX-MXFP8
- Reference
- not stated
- Eval data
- not stated (validation prompts, longest 73 prompt tokens)
- Context
- 73 prompt tokens longest validated
- Tool
- mlx-vlm / mlx-mtp, MXFP8 8-bit group 32, no calibration
Card snapshot main · updated 2026-08-15
Chungulus/Qwen3.8-27B-MLX-oQ4e-MTP
- Reference
- not stated (baseline vs candidate validation prompts)
- Eval data
- oMLX oQe packaged multilingual/code/tool/reasoning calibration corpus; short validation prompts (longest 73 tokens)
- Context
- 73 prompt tokens longest
- Tool
- oMLX / vllm-mlx native MTP
Card snapshot main · updated 2026-08-16
Chungulus/Qwen3.8-27B-MLX-oQ5
- Reference
- BF16 fixed-logit comparison
- Eval data
- fixed original text, 106 positions (not a public benchmark)
- Context
- short; longest validation prompt 73 tokens
- Tool
- oMLX / MLX-VLM logit scorer
Card snapshot main · updated 2026-08-16
Chungulus/Qwen3.8-27B-MLX-oQ8
- Reference
- BF16 log-probabilities of Qwen3.8-27B (stored float16 after float32 log-softmax)
- Eval data
- fixed original text, 106 positions, not a public benchmark
- Context
- longest validation prompt 73 tokens
- Tool
- MLX-VLM logit scorer / oMLX validation
Card snapshot main · updated 2026-08-16
cloudnathan5/Qwen3.8-27B-NVFP4a16-GPTQ
- Reference
- Qwen/Qwen3.8-27B BF16 measured through vLLM with identical settings
- Eval data
- wikitext-2-raw-v1 (test), 48 non-overlapping 4096-token windows, 196,560 tokens scored
- Context
- 4096
- Tool
- GPTQ NVFP4 weights with BF16 activations; vLLM 0.27.1
Card snapshot main · updated 2026-08-14
cloudnathan5/Qwen3.8-27B-NVFP4a4-FP8-GPTQ
- Reference
- Qwen/Qwen3.8-27B (BF16)
- Eval data
- wikitext-2-raw-v1 test, 48 non-overlapping 4096-token windows (196,560 tokens scored)
- Context
- 4096-token windows
- Tool
- vLLM with identical settings for both rows
Card snapshot main · updated 2026-08-14
cloudnathan5/Qwen3.8-27B-NVFP4a4-GPTQ
- Reference
- Qwen/Qwen3.8-27B (BF16), same vLLM settings
- Eval data
- wikitext-2-raw-v1 (test), 48 non-overlapping windows, 196,560 tokens scored
- Context
- 4096-token windows
- Tool
- token-level perplexity through vLLM
Card snapshot main · updated 2026-08-14
CodeMasterCody3D/taardis-27b-full-ternary
- Reference
- not stated (card says a comparison against the FP16 teacher is outstanding [L224-225])
- Eval data
- wikitext test (wiki.test.raw)
- Context
- c512, 274 chunks
- Tool
- llama-perplexity (TAARDIS llama.cpp fork)
Card snapshot main · updated 2026-09-27
cometkim/Qwen3.8-27B-nvfp4full-NInfer
- Reference
- task benchmarks (no KLD/PPL); gated against official artifact on GPQA-Diamond [L304]
- Eval data
- GPQA-Diamond, AIME 2026, LongBench v2
- Context
- native 262,144 (INT8 group-64 KV); 524,288 / 786,432 with YaRN
- Tool
- NInfer v3 fork engine, RTX 5090, thinking, temp 0.6 top_p 0.95 top_k 20, 3 seed rounds
Card snapshot main · updated 2026-09-16
curvedinf/Qwen3.8-27B-GPTQ-INT8-W8A8-GS128
- Reference
- gs32 reference checkpoint (not BF16)
- Eval data
- not stated
- Context
- not stated
- Tool
- teacher-forced KL-divergence sweep over 6 precision configurations
Card snapshot main · updated 2026-08-26
curvedinf/Qwen3.8-27B-GPTQ-INT8-W8A8-GS128-PTQR
- Reference
- BF16 teacher / baseline
- Eval data
- not stated
- Context
- not stated
- Tool
- teacher-forced KL-divergence probe at the serving gate (gate 0.02)
Card snapshot main · updated 2026-09-15
CySpiegel/Qwen3.8-27B-Int4-AutoRound
- Reference
- FP8 online per-tensor W8A16 build of the same base model
- Eval data
- HumanEval, HumanEval+, GSM8K 5-shot (raw-completions harness)
- Context
- not stated
- Tool
- vLLM, greedy decoding, thinking disabled
Card snapshot main · updated 2026-08-26
dancinlab/Qwen3.8-27B-Uncensored-GGUF
- Reference
- official Qwen/Qwen3.8-27B-FP8 (Base FP8)
- Eval data
- MMLU 0-shot n=300, MMLU-Pro n=250, GSM8K n=150, CMMLU n=500; imatrix on English+Chinese text
- Context
- not stated
- Tool
- vLLM serving of the abliterated build; llama.cpp for GGUFs
Card snapshot main · updated 2026-08-21
Davidmg0815/Qwen3.8-27B-MTP-GGUF
- Reference
- BF16 variant of Qwen3.8-27B (no external reference logits; pass-rate comparison across variants)
- Eval data
- HumanEval+ (164 tasks, execution-checked), MATH-500 (150 tasks)
- Context
- 16K-97K context; MATH-500 24k token budget headline
- Tool
- llama-server with custom bench scripts (bench_humaneval.py, bench_mathe.py)
Card snapshot main · updated 2026-09-27
dbirks/Qwen3.8-27B-NVFP4-AutoRound
- Reference
- BF16 base Qwen3.8-27B
- Eval data
- GSM8K full 1319, HumanEval pass@1 instruct, MMLU-Pro 100 questions per subject
- Context
- not stated
- Tool
- EleutherAI lm-evaluation-harness, thinking mode, temp 1.0, top-p 0.95, top-k 20
Card snapshot main · updated 2026-08-15
dbirks/Qwen3.8-27B-W4A16-AutoRound
- Reference
- BF16 base
- Eval data
- GSM8K full 1319 thinking; HumanEval pass@1 instruct; MMLU-Pro 100 questions/subject thinking; quantization calibration NeelNanda/pile-10k, 128 samples
- Context
- seqlen 2048 for calibration
- Tool
- EleutherAI lm-evaluation-harness; AutoRound W4A16
Card snapshot main · updated 2026-08-15
ddalcu/Qwen3.8-27B-MLX-Serve-6bit
- Reference
- bf16
- Eval data
- 64 held-out windows x 32 sampled positions (2048 distributions), no imatrix calibration for this build
- Context
- not stated
- Tool
- MLX affine quantization, group size 64; mlx-serve
Card snapshot main · updated 2026-09-29
ddalcu/Qwen3.8-27B-MLX-Serve-iQ-MLX-3.8bpw
- Reference
- bf16 source
- Eval data
- 32 held-out windows never seen in calibration corpus, 16 sampled positions each (512 positions)
- Context
- not stated
- Tool
- llmprobe (decode bench); top-1/KL tool not named
Card snapshot main · updated 2026-08-19
dealignai/Bonsai-2-27B-1bit-CRACK-GGUF
- Reference
- base PTQ1_0 ternary quant (prism-ml/Ternary-Bonsai-2-27B-gguf)
- Eval data
- MMLU n=2,280 (40 questions x 57 subjects, next-token letter-logit); HarmBench-320 refusal
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-09-18
dealignai/Bonsai-2-27B-Ternary-CRACK-GGUF
- Reference
- base ternary PQ2_0 quant (PrismML Bonsai 2 27B)
- Eval data
- MMLU n=2,280 (40 questions x 57 subjects), next-token letter-logit; HarmBench-320 for refusal
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-09-18
dealignai/Qwen3.8-27B-JANG_2D-CRACK
- Reference
- base vs CRACK on identical harness (benchmark deltas, no KL/PPL)
- Eval data
- MMLU 57-subject logit mode; HarmBench harm-240
- Context
- not stated
- Tool
- MLX runtime
Card snapshot main · updated 2026-09-08
dealignai/Qwen3.8-27B-JANG_4D-CRACK
- Reference
- unmodified base model on the identical harness (MMLU base vs CRACK)
- Eval data
- MMLU 57-subject logit mode; HarmBench harm-240 thinking-off
- Context
- not stated
- Tool
- MLX runtime
Card snapshot main · updated 2026-09-08
dealignai/Qwen3.8-27B-JANG_6D-CRACK
- Reference
- Base (unmodified model at the same quant level), identical harness
- Eval data
- MMLU 57-subject (logit mode); HarmBench harm-240
- Context
- not stated
- Tool
- MLX runtime; strict code/chemistry-aware classifier for HarmBench
Card snapshot main · updated 2026-09-08
dealignai/Qwen3.8-27B-MXFP8-CRACK
- Reference
- base model (unmodified) on identical harness
- Eval data
- MMLU 57-subject (logit mode); HarmBench harm-240
- Context
- not stated
- Tool
- MLX runtime
Card snapshot main · updated 2026-09-04
Derpyhue/Qwen3.8-27B-abliterated-2.55bpw-SC-SH
- Reference
- abliterated BF16 (Huihui-Qwen3.8-27B-abliterated), not the original Qwen model
- Eval data
- not stated for measurement; calibration traces from this model, 217 rows x 2048 columns
- Context
- not stated
- Tool
- EXL3 with sc_optimize; KL measured against abliterated BF16
Card snapshot main · updated 2026-10-14
Derpyhue/Qwen3.8-27B-abliterated-2.60bpw-SC-SH
- Reference
- Abliterated BF16 (Huihui-Qwen3.8-27B-abliterated)
- Eval data
- not stated for the KLD/PPL eval; quantization calibration used traces from this model (217 rows x 2048 columns)
- Context
- not stated
- Tool
- not stated (EXL3 quilter with sc_optimize, scaled-Hessian)
Card snapshot main · updated 2026-10-14
Derpyhue/Qwen3.8-27B-abliterated-2.65bpw-SC-SH
- Reference
- abliterated BF16 (Huihui-Qwen3.8-27B-abliterated), not the stock Qwen model
- Eval data
- not stated (calibration traces 217 rows x 2048 columns from this model)
- Context
- not stated
- Tool
- ExLlamaV3 sc_optimize with scaled-Hessian (-sh)
Card snapshot main · updated 2026-10-14
Derpyhue/Qwen3.8-27B-exl3-2.65bpw-SC
- Reference
- BF16 original
- Eval data
- not stated
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-10-12
Derpyhue/Qwen3.8-27B-exl3-2.65bpw-SC-SH
- Reference
- BF16 original
- Eval data
- not stated
- Context
- not stated
- Tool
- not stated (EXL3 quilter sc_optimize, scaled-Hessian -sh)
Card snapshot main · updated 2026-10-12
Derpyhue/Qwen3.8-27B-exl3-2.67bpw-SC-SH
- Reference
- BF16 original
- Eval data
- not stated
- Context
- not stated
- Tool
- not stated (EXL3; sc_optimize used for allocation)
Card snapshot main · updated 2026-10-12
Derpyhue/Qwen3.8-27B-exl3-2.70bpw-SC
- Reference
- BF16 original
- Eval data
- not stated
- Context
- not stated
- Tool
- not stated (sc_optimize allocation)
Card snapshot main · updated 2026-10-12
Derpyhue/Qwen3.8-27B-exl3-2.70bpw-SC-SH
- Reference
- BF16 original
- Eval data
- not stated
- Context
- not stated
- Tool
- not stated (EXL3 quilter with sc_optimize, -sh scaled Hessian)
Card snapshot main · updated 2026-10-12
Derpyhue/Qwen3.8-27B-exl3-2.75bpw-SC
- Reference
- BF16 original
- Eval data
- not stated
- Context
- not stated
- Tool
- not stated (sc_optimize used for bit allocation)
Card snapshot main · updated 2026-10-12
Derpyhue/Qwen3.8-27B-exl3-2.75bpw-SC-SH
- Reference
- BF16 original
- Eval data
- not stated
- Context
- not stated
- Tool
- EXL3 quilter with sc_optimize, scaled-Hessian (-sh)
Card snapshot main · updated 2026-10-12
Derpyhue/Qwen3.8-27B-exl3-2.80bpw-SC
- Reference
- BF16 original
- Eval data
- not stated
- Context
- not stated
- Tool
- EXL3 with sc_optimize noise-attribution bitrate allocation
Card snapshot main · updated 2026-10-12
Derpyhue/Qwen3.8-27B-exl3-2.80bpw-SC-SH
- Reference
- BF16 original
- Eval data
- not stated
- Context
- not stated
- Tool
- not stated (EXL3 KLD/perplexity eval)
Card snapshot main · updated 2026-10-12
deucebucket/Qwen3.8-27B-Cerebellum-GGUF
- Reference
- none (benchmarks, self-reported)
- Eval data
- ARC-Challenge, HellaSwag, MMLU-Redux, HumanEval/HumanEval+, perplexity 145 chunks
- Context
- 512 for perplexity; server -c 24576
- Tool
- llama-server -ngl 99 --parallel 4 --jinja
Card snapshot main · updated 2026-09-01
dfed24/Qwen3.8-27B-GPTQ-GGUF
- Reference
- bf16 model (PyTorch evaluator, L29-30); unsloth Q8_0 file for the llama.cpp check (L38)
- Eval data
- held-out wiki, code, multilingual sets
- Context
- 6 windows of 2,048 tokens per set
- Tool
- PyTorch evaluator; llama-perplexity --kl-divergence (CPU build)
Card snapshot main · updated 2026-10-13
drawthingsai/Qwen3.8-27B-GGUF
- Reference
- Qwen3.8-27B-BF16 benchmark scores
- Eval data
- AIME 2026, GPQA Diamond, IFEval (prompt_level_strict), BFCL v4 1K (1,002 examples, no Web Search)
- Context
- not stated
- Tool
- drawthingsai/benchmarks runner with evalscope 1.11.0 and llama-server, profile qwen3.8-thinking.json
Card snapshot main · updated 2026-09-13
DuoNeural/Qwen3.8-27B-TAP-DPQ-v2-GGUF
- Reference
- BF16 uncompressed model (Vanilla_BF16 reference)
- Eval data
- Wikitext-2 holdout perplexity; GSM8K N=100, HumanEval N=50, tool calling N=15; academic suite N=40 per task; imatrix calibration CodeInfused-v2
- Context
- not stated
- Tool
- llama.cpp CUDA backend on NVIDIA A100 80GB PCIe
Card snapshot main · updated 2026-09-30
eaddario/Qwen3.8-27B-GGUF
- Reference
- F16 GGUF baseline (logits saved)
- Eval data
- wikitext-2-raw-v1 (PPL/KLD); ARC, GPQA-Diamond, HellaSwag, MMLU-Redux, Truthful QA, WinoGrande 750 tasks each
- Context
- 1024 tokens for benchmark scores; PPL ctx not stated
- Tool
- llama-perplexity
Card snapshot main · updated 2026-08-23
EliovpAI/Qwen3.8-27B-W3Rot-INT3-Paiton-RDNA4
- Reference
- bf16 Qwen3.8-27B (KL(bf16 || 3-bit) of next-token distribution)
- Eval data
- 115 held-out chats, code, etc.; 22 long documents 8K-32K; 5 documents of 131K; 1,400-question MMLU-Pro subset
- Context
- up to 131K-token documents
- Tool
- Paiton image / vLLM plugin; MI355X reproduction and R9700 served
Card snapshot main · updated 2026-10-12
empero-ai/Qwen3.8-27B-Ridge-GGUF
- Reference
- BF16 GGUF converted by the author from the official checkpoint
- Eval data
- wiki-style calibration file, 80 chunks
- Context
- -c 512 -b 512
- Tool
- llama-perplexity
Card snapshot main · updated 2026-08-15
EschaLabs/Qwen3.8-27B-Escha-W2
- Reference
- same base model served as FP8 on same SGLang fork
- Eval data
- Commonsense-6 (full sets, thinking off); GPQA-Diamond n=198 and LiveCodeBench v6 2025 cut n=182 (thinking on, 28k budget)
- Context
- not stated
- Tool
- SGLang fork runtime, lm-eval style benchmarks
Card snapshot main · updated 2026-09-02
fasolack/Qwen3.8-27b-W4A16
- Reference
- dbirks/Qwen3.8-27B-W4A16-AutoRound reference build numbers (author did not measure own BF16 baseline)
- Eval data
- GSM8K, HumanEval (pass@1), MMLU-Pro (100/subject)
- Context
- GSM8K and MMLU-Pro 5-shot fewshot-as-multiturn; HumanEval temperature=1.0 n=1 and greedy
- Tool
- EleutherAI lm-evaluation-harness against vLLM served endpoint
Card snapshot main · updated 2026-09-16
FedericoSciuca/Qwen3.8-27B-depthaware-GGUF
- Reference
- Q8_0 reference build (PPL baseline)
- Eval data
- held-out WikiText-2, 32 chunks
- Context
- not stated
- Tool
- quantprobe / llama.cpp perplexity
Card snapshot main · updated 2026-10-07
finalvent/OddEye-0.1-27B-NVFP4
- Reference
- Original Qwen3.8-27B BF16 and a publicly distributed generic NVFP4 build, same settings incl. MTP speculative decoding
- Eval data
- 100 Japanese street photos (40 unseen), ~15 AI-written reference facts each; graded blind by a Claude model
- Context
- not stated
- Tool
- vLLM; custom michiyomi blind evaluation with prompt v5.2c, 8,192-token output cap
Card snapshot main · updated 2026-10-03
fraQtl/Qwen3.8-27B-Hi-Fi-GGUF
- Reference
- Q8_0 teacher quantized from the original pinned bf16 checkpoint
- Eval data
- locked code/math and general slices (pinned eval slice)
- Context
- not stated
- Tool
- symmetric top-20 KLD on the teacher support, 3 independent runs, byte-deterministic rig; GSM8K n=200 and MATH-500 n=500 0-shot CoT
Card snapshot main · updated 2026-08-21
Freaksterz/Qwen3.8-27B-SmoothQuant-W8A8-INT8
- Reference
- BF16 captured offline with transformers
- Eval data
- WikiText-2, 240 x 512-token windows = 122,640 positions
- Context
- 512-token windows
- Tool
- teacher-forced full-vocab KLD (top-K 512 with tail bound), candidates served through vLLM; floor-free paired estimator
Card snapshot main · updated 2026-09-04
Frosty40/Qwen3.8-27B-GGUF
- Reference
- full bf16 model
- Eval data
- three corpora: wikitext-2 prose, real code, real agent transcripts
- Context
- 512-token windows x 8
- Tool
- KL-divergence on llama.cpp (Intel Arc Pro B70)
Card snapshot main · updated 2026-08-18
GestaltLabs/Qwen3.8-27B-EXL3-11.5GB
- Reference
- Official FP8 Qwen3.8-27B
- Eval data
- 8 frozen prompt-prefix positions per task, 128 tasks (64 dev, 64 acceptance); task accuracy on same 128 tasks
- Context
- not stated
- Tool
- full-vocabulary natural-log KL (248,320 columns); greedy argmax decoding, non-thinking, max_new_tokens=512, Docker sandbox
Card snapshot main · updated 2026-09-09
gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090
- Reference
- not stated (head-to-head vs other NVFP4 builds, same harness)
- Eval data
- GPQA Diamond (198), AIME 2025 (30), MMLU-Pro (100), AA-LCR (100 q @ ~95K tok)
- Context
- 32,768-token budget
- Tool
- NVIDIA Model Optimizer; vLLM on one RTX 5090, temp 1.0, top_k 20, top_p 0.95
Card snapshot main · updated 2026-09-10
glyd/Qwen3.8-27B-kestrel
- Reference
- bf16 original (commit 1d4bf0f2)
- Eval data
- WikiText-2 (KL); GSM8K 10-shot, 1319 questions, non-thinking
- Context
- not stated
- Tool
- glyd serve with Inspect AI 0.3.278 (inspect_evals 0.24.0); bf16 served by vLLM
Card snapshot main · updated 2026-10-10
glyd/Qwen3.8-27B-penguin
- Reference
- bf16 original
- Eval data
- WikiText-2
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-10-10
glyd/Qwen3.8-27B-swift
- Reference
- bf16 original
- Eval data
- WikiText-2 (KL); GSM8K 10-shot 1319 questions non-thinking
- Context
- not stated
- Tool
- glyd serve with Inspect AI 0.3.278 / inspect_evals 0.24.0; bf16 served by vLLM
Card snapshot main · updated 2026-10-10
gratex/Qwen3.8-27B-W4A16-g128-sym-GPTQ
- Reference
- BF16 base
- Eval data
- wikitext-2-raw-v1 test, 100 non-overlapping chunks (24,426 scored tokens); downstream via lm-evaluation-harness v0.4.12
- Context
- 512 tokens per chunk
- Tool
- vLLM completions API with logprobs; KLD via sparse top-20 logprob approximation
Card snapshot main · updated 2026-09-03
greghavens/Qwen3.8-27B-bnb-4bit
- Reference
- bf16 base, paired (identical prompts, order, greedy, pinned batch size)
- Eval data
- wikitext-2; IFEval 541 prompts; BFCL v4 all 13 single-turn categories 3641 prompts; NoveltyBench NB-Curated 100 prompts k=10
- Context
- 4096-token windows for perplexity
- Tool
- bitsandbytes NF4 (data-free), same harness for bf16; thinking disabled
Card snapshot main · updated 2026-09-04
gubernac/qwen3.8-27b-unc-fp8
- Reference
- official Qwen/Qwen3.8-27B-FP8 served identically, same scripts and settings
- Eval data
- MMLU (300), MMLU-Pro CoT (250), GSM8K CoT (150), CMMLU (500); WikiText-2-raw PPL
- Context
- not stated
- Tool
- vLLM, block-FP8 + FP8 KV + MTP
Card snapshot main · updated 2026-08-16
guideboardlabs/Qwen3.8-27B-Q3-DOWN-XS-GGUF
- Reference
- UD-IQ3_XXS 3-bit control (no bf16 anchor built)
- Eval data
- wikitext-2-raw-v1 test set
- Context
- 8 x 512 chunks, batch 128
- Tool
- llama.cpp llama-perplexity, Vulkan, RX 5700 XT
Card snapshot main · updated 2026-08-21
HangGlidersRule/Darkstar-Qwen3.8-27B-Base-ModelOpt-W4A16-NVFP4-Mixed-FP8
- Reference
- upstream BF16 control
- Eval data
- GPQA Diamond (198), thinking off; harmful-prompt (200) and safe (83) refusal sets
- Context
- not stated for eval (runtime context 126,144)
- Tool
- vLLM 0.27.1; calibration cnn_dailymail + nemotron-post-training-dataset-v2
Card snapshot main · updated 2026-08-21
Hent-AI/Qwen3.8-27B-EXL3-1.14514bpw
- Reference
- not stated (no bf16 comparison; absolute scores only)
- Eval data
- 6 benchmarks (MMLU-Redux, MuSR, GSM8K, HumanEval+, IFEval, BFCL v3) plus custom 0-shot Japanese set of 8,245 questions (JCommonsenseQA, JNLI, JSQuAD, MGSM-ja)
- Context
- not stated
- Tool
- EvalScope 1.12.0 / ExLlamaV3 1.5.3
Card snapshot main · updated 2026-10-06
HermiHg/Qwen3.8-27B-mmproj-Q5_K-MIX-GGUF
- Reference
- BF16 mmproj projector
- Eval data
- 11 vision benchmarks, 5,251 paired documents (TextVQA, OCRBench-v2, GQA, ScienceQA, MMStar, ChartQA, five BLINK categories)
- Context
- not stated
- Tool
- llama-server + lmms-eval, temperature 0, 24 GB NVIDIA GPU
Card snapshot main · updated 2026-09-01
HivenetQuant/Qwen3.8-27B-NVFP4
- Reference
- original FP16 weights, paired head-to-head, identical prompts/harness/decoding
- Eval data
- mmlu_pro, gpqa_diamond, aime25, ifeval, mmmu_pro, ai2d, livecodebench v6, bfcl, niah, mtp accept rate
- Context
- 256k max context, FP8 KV cache
- Tool
- vLLM serve config in card; harness details in REPRODUCIBILITY.md
Card snapshot main · updated 2026-08-25
huginnfork/Qwen3.8-27B-FP8
- Reference
- bf16 Qwen/Qwen3.8-27B
- Eval data
- KLD: 8 samples from neuralmagic/calibration (LLM split); PPL: wikitext-2-raw test
- Context
- KLD max_seq=1024; PPL non-overlapping 2048-token chunks
- Tool
- not stated
Card snapshot main · updated 2026-08-14
huginnfork/Qwen3.8-27B-NVFP4A16
- Reference
- bf16 Qwen/Qwen3.8-27B
- Eval data
- KLD: 8 samples from neuralmagic/calibration (LLM split); PPL: wikitext-2-raw test
- Context
- KLD max_seq=1024; PPL non-overlapping 2048-token chunks
- Tool
- custom harness, transformers
Card snapshot main · updated 2026-08-14
ilintar/qwen3.8-27b-gguf-strix-halo
- Reference
- source BF16 GGUF
- Eval data
- 16 held-out chunks
- Context
- 512-token chunks
- Tool
- perplexity (small local evaluation, not a general quality benchmark)
Card snapshot main · updated 2026-09-04
inferencerlabs/Qwen3.8-27B-MLX-Q9
- Reference
- Base (unquantized Qwen3.8-27B)
- Eval data
- not stated
- Context
- not stated
- Tool
- Inferencer app on M3 Ultra, modified MLX
Card snapshot main · updated 2026-08-14
intel-ai/Qwen3.8-27B-MXFP8-CT-AutoRound
- Reference
- BF16 Qwen/Qwen3.8-27B upstream-reported scores (not re-run)
- Eval data
- gsm8k 5-shot (chat template), piqa, mmlu, hellaswag 0-shot
- Context
- max_model_len 8192
- Tool
- lm-eval 0.4.13 with vllm backend, seed 42, single B300
Card snapshot main · updated 2026-09-21
intel-ai/Qwen3.8-27B-MXFP8-MXFP8MTP-CT-AutoRound
- Reference
- BF16 Qwen/Qwen3.8-27B (upstream-reported table, not re-run)
- Eval data
- lm-eval gsm8k 5-shot, piqa, mmlu, hellaswag 0-shot
- Context
- max_model_len 8192
- Tool
- lm-eval with vLLM 0.22.1 backend, seed 42, single B200, text tower only, thinking off
Card snapshot main · updated 2026-10-12
ionos/Qwen3.8-27B-GPTQ-Int4
- Reference
- unquantized base (BF16)
- Eval data
- OmniDocBench, MathVista, MMMU, V*Bench, MMLU-Pro, MMLU-DE, BFCL, IFEval
- Context
- not stated
- Tool
- paired comparisons, same prompts and settings, greedy decoding
Card snapshot main · updated 2026-09-24
ISTA-DASLab/Qwen3.8-27B-3Bit-GSQ
- Reference
- original Qwen/Qwen3.8-27B
- Eval data
- AIME 2025, GPQA Diamond
- Context
- not stated
- Tool
- xhigh thinking enabled for both models
Card snapshot main · updated 2026-08-20
IsValorum/Qwen3.8-27B-VAL-APEX-I-MiniPlus-V3-GGUF
- Reference
- unmodified BF16 Base (PPL 5.7942)
- Eval data
- WikiText-2
- Context
- not stated
- Tool
- llama.cpp perplexity; imatrix 1,000 chunks x 4,096 tokens
Card snapshot main · updated 2026-10-16
IsValorum/Qwen3.8-27B-VAL-APEX-I-NanoPlus-V3-GGUF
- Reference
- unmodified BF16 Base (PPL 5.7942)
- Eval data
- WikiText-2
- Context
- not stated
- Tool
- not stated (certified perplexity evaluation)
Card snapshot main · updated 2026-10-16
jaakkomoller/Qwen3.8-27B-q3mix-NInfer
- Reference
- pure-q4 profile of same model (head-to-head); stored vLLM reference run
- Eval data
- LiveCodeBench v6 175-problem subset; small quality gate (35 items)
- Context
- greedy, max 2,000 tokens, single sample
- Tool
- NInfer
Card snapshot main · updated 2026-09-27
Jackrong/Qwopus3.8-27B-Flash-GGUF
- Reference
- Qwen3.8 base comparison (same Q5_K_M + MTP configuration)
- Eval data
- MMLU-Pro mixed set, 2,500 questions
- Context
- not stated
- Tool
- author-provided local results
Card snapshot main · updated 2026-09-10
JANGQ-AI/Qwen3.8-27B-JANG_4D
- Reference
- bf16 model
- Eval data
- held-out text set by topic plus own bf16 reasoning traces (efforts low/medium/xhigh)
- Context
- not stated
- Tool
- importance-matrix weighted fit, AWQ folded into norms, GPTQ on MLX grid; KL and top-1 vs bf16
Card snapshot main · updated 2026-10-06
JANGQ-AI/Qwen3.8-27B-JANGH2
- Reference
- bf16 model
- Eval data
- held-out text set by topic; own bf16 reasoning traces (thinking + answer tokens)
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-10-06
jaromer/0bserverx-Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF
- Reference
- base Qwen3.8-27B
- Eval data
- 100 harmful-behaviors prompts (refusals); KL dataset not stated
- Context
- not stated
- Tool
- not stated (heretic ARA abliteration; KL vs base)
Card snapshot main · updated 2026-08-16
jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge
- Reference
- BF16 Qwen3.8-27B
- Eval data
- not stated; also locked 20-task HumanEval/EvalPlus panel
- Context
- not stated
- Tool
- local matched measurements on one Ryzen AI Max+ 395 / Radeon 8060S
Card snapshot main · updated 2026-08-19
jcbtc/Qwen3.8-27B-IU4-Kairic-Edge
- Reference
- not stated (pass counts on tasks; no BF16 reference metric)
- Eval data
- 164-task chat-adapted HumanEval/EvalPlus harness
- Context
- Kairic 262,144 context with 8 GiB prompt cache; comparison runs 65,536 context
- Tool
- EvalPlus-based chat-adapted harness, deterministic sampling, one trajectory per task, no repair retry
Card snapshot main · updated 2026-09-24
jclyons52/Qwen3.8-27B-Unsloth-Dynamic-v3-mlx
- Reference
- not stated (no bf16 anchor; perplexity only)
- Eval data
- wikitext-2, 64 x 512-token samples, seed 123
- Context
- 512
- Tool
- not stated; M1 Pro 32 GB
Card snapshot main · updated 2026-08-25
jdominick05/Qwen3.8-27B-24GB-262k-GGUF
- Reference
- Q8_0 of the same BF16 weights
- Eval data
- technical text (decompiler output, C, notes) interleaved with coding-agent transcripts
- Context
- 65,536 tokens (16 x 4,096)
- Tool
- llama-perplexity --kl-divergence
Card snapshot main · updated 2026-10-13
jialinyyzz/Qwen3.8-27B-abliterated-MLX-4bit
- Reference
- original Qwen3.8-27B (bf16 master, unabliterated)
- Eval data
- mlabonne/harmful_behaviors test[:100] (refusals); ARC-Challenge, HellaSwag, Winogrande, OpenBookQA, GSM8K CoT
- Context
- not stated
- Tool
- Heretic keyword-refusal scorer; lm-eval-harness loglikelihood, GSM8K generative CoT; measured on bf16 master, not the 4-bit file
Card snapshot main · updated 2026-08-16
jialinyyzz/Qwen3.8-27B-abliterated-MLX-8bit
- Reference
- original Qwen3.8-27B (pre-abliteration), harmless-prompt first-token distribution for KL
- Eval data
- ARC-Challenge, HellaSwag, Winogrande, OpenBookQA (loglikelihood), GSM8K CoT generative; refusals on mlabonne/harmful_behaviors test[:100]
- Context
- not stated
- Tool
- lm-eval-harness; Heretic refusal scorer
Card snapshot main · updated 2026-08-16
Jianqiao1/Qwen3.8-27b-MoQ-GGUF
- Reference
- BF16 logits from original Qwen3.8-27B GGUF
- Eval data
- WikiText-2 wiki.test.raw
- Context
- 512
- Tool
- llama.cpp build 10276 (commit 6ea215d17), CUDA, RTX 5090
Card snapshot main · updated 2026-08-17
jkim96/Qwen3.8-27B-DASHQ-Q2-GGUF
- Reference
- none (absolute perplexity, no bf16 row)
- Eval data
- WikiText-2 test; C4 validation (256 x 2048 tokens)
- Context
- 2048
- Tool
- llama-perplexity
Card snapshot main · updated 2026-10-06
JMVRoill/Qwen3.8-27B-Uncensored-nvfp4-NInfer
- Reference
- orcarouter FP8 build / base FP8 (not this artifact)
- Eval data
- MMLU, MMLU-Pro, GSM8K, CMMLU, XSTest-safe (sizes not stated)
- Context
- not stated
- Tool
- orcarouter's measurements; NInfer converter
Card snapshot main · updated 2026-09-09
Jon-Nielsen/Qwen3.8-27B-exl3-4.00bpw
- Reference
- not stated (absolute perplexity only)
- Eval data
- PPL (40K tech)
- Context
- 16K-token prompt / 512 gen, 262K context config, Q6 KV
- Tool
- ExLlamaV3 EXL3, RTX PRO 6000, vision disabled, MTP draft
Card snapshot main · updated 2026-08-14
Jon-Nielsen/Qwen3.8-27B-exl3-6.00bpw
- Reference
- not stated
- Eval data
- 40K tech text (PPL)
- Context
- 16K-token prompt / 512 gen, Q6 KV, 262K context config
- Tool
- ExLlamaV3 on RTX PRO 6000
Card snapshot main · updated 2026-08-14
JRamirez-UAB/Qwen3.8-27B-EXL3-3.85bpw-24GB
- Reference
- BF16 Qwen3.8-27B
- Eval data
- internal sysadmin/agent eval set (unpublished) and turboderp's public qbench_prompts_gen.json
- Context
- not stated
- Tool
- exllamav3 qbench (GPTQ evaluated as BF16 copy of dequantized weights)
Card snapshot main · updated 2026-10-16
JRamirez-UAB/Qwen3.8-27B-GPTQ-W4A16-embed-int4-24GB
- Reference
- BF16 Qwen3.8-27B
- Eval data
- internal sysadmin/agent eval set (unpublished) and turboderp's public qbench_prompts_gen.json
- Context
- not stated
- Tool
- exllamav3 qbench (GPTQ evaluated as BF16 copy of dequantized weights)
Card snapshot main · updated 2026-10-14
JRamirez-UAB/Qwen3.8-27B-GPTQ-W4A16-g64-embed-int8-24GB
- Reference
- BF16 Qwen3.8-27B
- Eval data
- internal sysadmin/agent set (unpublished) and turboderp qbench_prompts_gen.json (public)
- Context
- not stated
- Tool
- exllamav3 qbench
Card snapshot main · updated 2026-10-14
julianmb/Qwen-3.8-27B-ROCmFP4-FAST-GGUF
- Reference
- not stated
- Eval data
- wikitext-2
- Context
- 512
- Tool
- llama-perplexity, ROCm0
Card snapshot main · updated 2026-08-22
juzbox/Qwen3.8-27B-NVFP4-GDN
- Reference
- BF16 original Qwen/Qwen3.8-27B
- Eval data
- KL: 35,328 positions, ja/en/c... (truncated in candidate); GSM8K 300, BFCL v3 1,000
- Context
- not stated
- Tool
- vLLM v0.29.0 and a custom inference engine
Card snapshot main · updated 2026-10-11
kacaforyah/Opti-27B
- Reference
- Qwen3.8-27B 16-bit original (F16 logits)
- Eval data
- wikitext-2 test (also C4, Python stdlib, Pride and Prejudice); MMLU-Pro 12,032 items
- Context
- 1,024-token windows, 96 windows (64 for paired KLD)
- Tool
- llama-perplexity (Opti runtime patch on llama.cpp)
Card snapshot main · updated 2026-09-30
kaushikvira/Qwen3.8-27B-nvfp4full-dflash2-NInfer
- Reference
- none
- Eval data
- IFBench prompt-strict, GSM8K-200
- Context
- not stated
- Tool
- not stated; single RTX 5090, NInfer engine
Card snapshot main · updated 2026-09-26
kaushikvira/Qwen3.8-27B-nvfp4full-dflash2-NInfer-v3
- Reference
- not stated
- Eval data
- IFBench prompt-strict and GSM8K-200 (family table); repo gate probes (needle, tool calls)
- Context
- 262,144 context for gate run
- Tool
- repo gate.sh and perf.py; llama-benchy for speed
Card snapshot main · updated 2026-09-26
kaushikvira/Qwen3.8-27B-swift15-nvfp4full-dflash2-NInfer-v3
- Reference
- no BF16 reference; same-session A/B vs the publisher's Swift-1.0-abliterated build, temp 0, no cache reuse
- Eval data
- IFBench (n=300), GSM8K-200; quantization calibration 512 Ultrachat samples
- Context
- 2048 (calibration seq)
- Tool
- llm-compressor NVFP4; NInfer engine
Card snapshot main · updated 2026-10-04
kaushikvira/Qwen3.8-27B-swift15-uncensored-ajgazin-nvfp4full-dflash2-NInfer-v3
- Reference
- Swift-1.5 (prod, censored) artifact, same session A/B
- Eval data
- IFBench prompt-strict (n=300), GSM8K-200, harmful_behaviors test[:100] refusal, harmless sanity n=50
- Context
- not stated
- Tool
- NInfer engine, temp 0, single RTX 5090
Card snapshot main · updated 2026-10-04
kaushikvira/Qwen3.8-27B-thinkingcap-nvfp4full-dflash2-NInfer-v3
- Reference
- Swift-1.5 production artifact (same engine build, same session A/B); upstream base Qwen3.8-27B for BottleCap's own eval
- Eval data
- GSM8K-200, IFBench (n=300)
- Context
- not stated
- Tool
- NInfer engine, temp 0, single RTX 5090, fresh generations, no cache reuse
Card snapshot main · updated 2026-10-04
kayceesamuel/Qwen3.8-27B-NF4DQ
- Reference
- BF16 PPL 6.9530 as the reference row
- Eval data
- wikitext-2, full test file
- Context
- 512
- Tool
- NF4DQ llama.cpp fork, A100-80GB
Card snapshot main · updated 2026-08-31
kdqemre/Qwen3.8-27B-RQ2_K_L-GGUF
- Reference
- Q5_K_M baseline (accuracy comparison, no BF16 reference)
- Eval data
- GSM8K-100, 8-shot, Exact Match
- Context
- not stated
- Tool
- greedy decode, MTP off, RTX 4090 power-limited to 150W; author's llama.cpp fork
Card snapshot main · updated 2026-08-22
kekzle/Qwen3.8-27B-NVFP4-vllm
- Reference
- none stated (absolute values)
- Eval data
- ppl_corpus_45k.txt (13.5k tokens); degen_suite.py 45 checks
- Context
- not stated
- Tool
- imp 0.27, RTX 5090, batch 1, greedy
Card snapshot main · updated 2026-08-31
kingjones777/Qwen3.8-27B-ROCmFP4-STRIX-MTP-GGUF
- Reference
- perplexity only, compared against Q4_K_M reference; no KLD
- Eval data
- held-out wikitext-2
- Context
- 2048, 40 chunks
- Tool
- llama.cpp (ROCmFPX fork)
Card snapshot main · updated 2026-09-01
kirillbilchenko/Qwen3.8-27B-MXFP4-MxWave
- Reference
- BF16 source Qwen/Qwen3.8-27B (pinned revision)
- Eval data
- WikiText-2 raw test: 316 windows of 4,096 characters, 297,199 scored tokens (PPL); 128 held-out WikiText contexts up to 512 tokens, full 248,320 vocab (KL); GSM8K 100-example 5-shot pilot
- Context
- up to 512 tokens for KL; 4,096-character windows for PPL
- Tool
- MxWave H64 block-Hessian MXFP4; calibration 64 sequences of 512 tokens from pile_val_test
Card snapshot main · updated 2026-09-16
kirillbilchenko/Qwen3.8-27B-MxWave-MXFP4-FP8
- Reference
- BF16 source
- Eval data
- WikiText-2 prompt PPL (316 windows of 4,096 characters, 297,199 scored tokens); untouched 96-context distribution holdout; GSM8K full 1,319 five-shot
- Context
- 4,096-character windows; GSM8K 1,024 max generated tokens, greedy
- Tool
- lm-evaluation-harness 0.4.13
Card snapshot main · updated 2026-09-19
komeijishiki/Qwen3.8-27B-EXL3-SC-3.40bpw-H6-V6
- Reference
- not stated (predicted KLD by recipe; IQ4_XS used as comparison)
- Eval data
- not stated
- Context
- not stated
- Tool
- turboderp ExLlamaV3 v1.4.4 Full SC recipe
Card snapshot main · updated 2026-08-31
lackonendes/PAW-125B-FLASH-NEXT-X3
- Reference
- none; absolute scores only, no BF16 comparison (KL vs BF16 parent explicitly unmeasured)
- Eval data
- MMLU-Pro 500 items, GPQA 198, IFBench-64 (64-item sentinel), HumanEval/HumanEval+ 164, MBPP/MBPP+ 378, GSM8K 100
- Context
- not stated
- Tool
- no-think fastgate harness, greedy, max_tokens per benchmark
Card snapshot main · updated 2026-10-03
lackonendes/PAW-27B-GGUF
- Reference
- none; peer comparison against two IQ2XXS GGUFs, no bf16 denominator
- Eval data
- MMLU-Pro (500), GSM8K (100), HumanEval+, MBPP+, IFBench-64 strict (64-item sentinel)
- Context
- not stated
- Tool
- byte-identical harnesses on same machine
Card snapshot main · updated 2026-10-02
lackonendes/PAW-27B-X3-GGUF
- Reference
- Q8 parent (author's own run); KL vs Q8 parent for X3.1 comparison
- Eval data
- MMLU-Pro /500, IFBench /64 sentinel, HumanEval /164, MBPP /378, GSM8K /100; KL on code (chat-framed and raw)
- Context
- not stated
- Tool
- llama-bench for speed; llama-paw build; one RTX 3090, byte-identical harnesses
Card snapshot main · updated 2026-10-02
lackonendes/PAW-27B-X3.1-GGUF
- Reference
- published BF16 Base column from the Qwen/Qwen3.8-27B card (not rerun by author)
- Eval data
- AIME 2026, GPQA-Diamond, MMLU-Pro, LiveCodeBench v6, IFBench
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-10-02
Launch80/Qwen3.8-27B-PARO-int5
- Reference
- bf16 served on the same stack
- Eval data
- wikitext, 96 x 500-char chunks (KL); GSM8K 500q greedy
- Context
- not stated
- Tool
- KL(bf16 || candidate) top-5 and top-256, 2 x R9700 TP=2, fp8 KV
Card snapshot main · updated 2026-09-12
LeaderboardModel1/Qwen3.8-27B-AutoRound-MXFP4-ModelFree
- Reference
- not stated
- Eval data
- hellaswag, piqa
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-08-17
LeaderboardModel1/Qwen3.8-27B-AutoRound-W4A16
- Reference
- not stated
- Eval data
- gsm8k, hellaswag, mmlu, piqa
- Context
- not stated
- Tool
- AutoRound W4A16
Card snapshot main · updated 2026-08-25
LeaderboardModel1/Qwen3.8-27B-AutoRound-W4A16-tuning
- Reference
- not stated
- Eval data
- gsm8k, hellaswag, mmlu, piqa
- Context
- not stated
- Tool
- AutoRound W4A16
Card snapshot main · updated 2026-08-25
LeaderboardModel1/Qwen3.8-27B-AutoRound-W4A16-v2
- Reference
- not stated
- Eval data
- not stated
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-08-26
leonsarmiento/Qwen3.8-27B-3bit-mlx
- Reference
- vanilla uniform 4-bit mlx-community/Qwen3.8-27B-4bit (no BF16 reference)
- Eval data
- MMLU and MMLU_PRO, 50-question samples each, instruct mode
- Context
- not stated
- Tool
- oMLX (Engine Auto, Context Code/Python)
Card snapshot main · updated 2026-08-18
leonsarmiento/Qwen3.8-27B-3bit-mtp-mlx
- Reference
- not stated
- Eval data
- MMLU 100/14042, MMLU_PRO 100/12032, HUMANEVAL 50/164, MBPP 50/500 (sampled)
- Context
- not stated
- Tool
- oMLX v0.6.1, instruct mode (thinking off)
Card snapshot main · updated 2026-09-09
libertyhhn/Qwen3.8-27B-GPTQ-W4A16-g128
- Reference
- Dense bf16 of same model, same pipeline
- Eval data
- wikitext-2 test (12 windows of 2,048 tokens); MMLU up to 512 questions per subject (12,233 total, no chat template); GSM8K full test 1,319, flexible-extract, chat template, thinking off
- Context
- 2,048-token windows for perplexity
- Tool
- vLLM 0.31.0 and lm-eval 0.4.13, bf16 activations
Card snapshot main · updated 2026-10-10
Litwein/Qwen3.8-27B-oQ4e-DWQ-MTP-Vision-MLX
- Reference
- oQ8e teacher made from the same base model (not BF16) for the held-out table; teacher itself is 0.019 KL vs bf16 base
- Eval data
- held-out calibration distribution (code/agent heavy mix, 62.5% agent + code)
- Context
- 512-token windows
- Tool
- KL over teacher's top-1024 logits at temperature 2.0 (DWQ objective); oMLX
Card snapshot main · updated 2026-08-23
lmcoleman/Qwen3.8-27B-MagicQuant-GGUF
- Reference
- BF16 baseline perplexity 6.7443
- Eval data
- wikitext-2
- Context
- 512, 100 chunks
- Tool
- perplexity measurement during MagicQuant search
Card snapshot main · updated 2026-08-20
lmcoleman/Qwen3.8-27B-ROCmFPX-GGUF
- Reference
- BF16 baseline PPL 6.7443
- Eval data
- wikitext-2
- Context
- 512, 100 chunks
- Tool
- ciru-ai/ROCmFPX llama.cpp fork, MagicQuant methodology
Card snapshot main · updated 2026-08-16
Lottolabs/Qwen3.8-27B-TT-Mixed-BFP4-BFP8-P150
- Reference
- two tables: TT all-BFP8 build (same runtime) and CPU BF16 reference
- Eval data
- 34-record eval set (14 multi-turn chats, 12 Python files, 8 instructions), 42,129 scored positions; teacher-forced on P150
- Context
- not stated for checkpoint tables; 32K-262K for KV-cache table
- Tool
- teacher-forced device runs on Tenstorrent P150
Card snapshot main · updated 2026-10-06
lribeiro/Qwen3.8-27B-FP8-Pessoa
- Reference
- BF16 Qwen3.8-27B (shared BF16 LM head replay)
- Eval data
- fidelity suite v3, 136 contexts, 278,392 scored positions (calibration: malaiwah/qwen38-27b-fidelity-suite-v3, 181 contexts)
- Context
- 2048 tokens
- Tool
- custom hidden-state KLD harness, two-pass log-sum-exp; llm-compressor GPTQ
Card snapshot main · updated 2026-08-21
Lucebox/Qwen3.8-27B-IQ4_XS-fast-GGUF
- Reference
- Q8_0 of Qwen3.8-27B
- Eval data
- wikitext-2 (300 x 512-token chunks) for KLD/PPL; HumanEval-164 and GSM8K first 200 through serving path
- Context
- 512-token chunks
- Tool
- llama.cpp paired KL; lucebox serving path with exact-verify speculation
Card snapshot main · updated 2026-09-22
lued/Qwen3.8-27B-INT8-W8A16-DFlash2
- Reference
- original BF16 model
- Eval data
- not stated (TECHNICAL.md holds details)
- Context
- 262K serving
- Tool
- not stated
Card snapshot main · updated 2026-08-21
lued/Qwen3.8-27B-INT8-W8A16-MTP
- Reference
- original BF16 Qwen/Qwen3.8-27B checkpoint (teacher)
- Eval data
- 4,563 teacher-forced positions over 8 short prompts plus a 4,096-token long-context probe (first tokens of the Qwen3.8 model card, 4,095 scored positions)
- Context
- short prompts; 4,096 tokens for long-context probe
- Tool
- custom KLD over full 248,320-token vocab, FP32 log_softmax, Hugging Face decompressed path, use_cache=False
Card snapshot main · updated 2026-08-15
machinez/d3-NVFP4
- Reference
- not stated (task scores of the d3 decision model; no comparison against an unquantized d3)
- Eval data
- Jev Decision Index 0.3.1 text board, vision board, public vision benchmarks, Perception Test
- Context
- not stated
- Tool
- official Jev 0.3 kit
Card snapshot main · updated 2026-10-11
magiccodingman/Qwen3.8-27B-MagicQuant-GGUF
- Reference
- not stated (KLD against source Qwen3.8-27B model)
- Eval data
- not stated
- Context
- not stated
- Tool
- MagicQuant benchmark flow (llama.cpp), generic MagicQuant imatrix; six files rebuilt with Unsloth imatrix
Card snapshot main · updated 2026-08-21
maglun/Qwen3.8-27B-MLX-Mixed-3.80bpw
- Reference
- BF16 text model (measured on a different machine, M4 Pro, for the 2nd table)
- Eval data
- held-out chat-framed corpus: 6 documents (launch table), 4 documents / 6,057 tokens (2026-10-15 table); HumanEval+ n=8, GSM8K n=12
- Context
- not stated
- Tool
- mlx 0.32.3 / mlx-vlm 0.7.6 / mlx-lm 0.32.0, greedy, reasoning_effort=low, 8192-token cap, Apple M5 Pro 64 GB
Card snapshot main · updated 2026-10-15
maglun/Qwen3.8-27B-MLX-Mixed-4.95bpw
- Reference
- BF16 text source
- Eval data
- held-out chat-framed corpus of 4 documents / 6,057 tokens (PPL); HumanEval+ n=8; GSM8K n=12
- Context
- not stated
- Tool
- mlx / mlx-vlm / mlx-lm, greedy decoding, reasoning_effort=low, 8,192-token cap
Card snapshot main · updated 2026-10-15
mahadev9/Qwen3.8-27B-fp8
- Reference
- bf16 baseline
- Eval data
- prose = WikiText-2; agentic = tool-calling turns (schema + query + call) rendered with chat template
- Context
- seq len 2048
- Tool
- not stated
Card snapshot main · updated 2026-09-05
majentik/Qwen3.8-27B-MLX-2bit
- Reference
- not stated
- Eval data
- arc_easy, hellaswag
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-08-15
majentik/Qwen3.8-27B-MLX-3bit
- Reference
- not stated
- Eval data
- arc_easy, hellaswag (sample counts not stated)
- Context
- not stated
- Tool
- mlx_lm.convert 0.31.3 (affine 3-bit, group 32)
Card snapshot main · updated 2026-08-15
majentik/Qwen3.8-27B-MLX-4bit
- Reference
- none stated
- Eval data
- arc_easy, hellaswag (sample size not stated)
- Context
- not stated
- Tool
- not stated (mlx-lm 0.31.3 used for conversion)
Card snapshot main · updated 2026-08-15
majentik/Qwen3.8-27B-MLX-5bit
- Reference
- none
- Eval data
- arc_easy, hellaswag
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-08-15
majentik/Qwen3.8-27B-MLX-6bit
- Reference
- not stated
- Eval data
- arc_easy and hellaswag (limited sample suggested by round scores)
- Context
- not stated
- Tool
- not stated (mlx_lm.convert for quantization)
Card snapshot main · updated 2026-08-15
majentik/Qwen3.8-27B-MLX-8bit
- Reference
- not stated
- Eval data
- arc_easy, hellaswag
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-08-15
majentik/Qwen3.8-27B-MLX-MXFP4
- Reference
- not stated
- Eval data
- arc_easy, hellaswag (sample size not stated)
- Context
- not stated
- Tool
- not stated (mlx-lm 0.31.3 used for conversion)
Card snapshot main · updated 2026-08-15
malaiwah/Qwen3.8-27B-EXL3-EDA-research
- Reference
- shared BF16 reference capture, shared BF16 lm_head
- Eval data
- shard 0 of v5 held-out suite, 512 contexts, 1,048,064 scored positions, 330 source clusters
- Context
- not stated
- Tool
- tools/fidelity.py replay (hidden state after final norm replayed through BF16 lm_head); paired per context
Card snapshot main · updated 2026-08-20
malaiwah/Qwen3.8-27B-EXL3-K5K6
- Reference
- BF16 reference (KL(BF16 reference || candidate), nats, full vocabulary, body-only, shared BF16 head replay)
- Eval data
- v5 10M suite: 5,120 contexts x 2,047 scored positions = 10,480,640 positions, 842 source clusters; shard-0 comparison uses 512 contexts / 1,048,064 positions
- Context
- 2,048-token contexts
- Tool
- vLLM capture + shared-BF16-head replay (tools/kld_ladder.sh, kld_aggregate.py); llama.cpp for GGUF rows
Card snapshot main · updated 2026-09-08
malaiwah/Qwen3.8-27B-EXL3-K5K6-context
- Reference
- BF16 Qwen3.8-27B, full-vocab KL(BF16 || candidate) in nats, body-only through one shared BF16 head
- Eval data
- v5 held-out suite: 5,120 contexts, 10,480,640 scored positions, 842 source clusters (also v3 136-context, v4 36-context, shard-0 512-context subsets)
- Context
- 2048 tokens per context
- Tool
- vLLM capture + shared-BF16-head replay; llama.cpp llama-perplexity for GGUF rows
Card snapshot main · updated 2026-09-08
malaiwah/Qwen3.8-27B-EXL3-K5K6-hydrated
- Reference
- BF16 Qwen/Qwen3.8-27B, full-vocabulary KL(BF16 reference || candidate) in nats, body-only, shared BF16 LM head replay
- Eval data
- v5 held-out suite: 5,120 contexts, 10,480,640 scored positions, 842 source clusters (shard 0 = 512 contexts, 1,048,064 positions)
- Context
- 2048 tokens per context
- Tool
- vLLM capture + replay pipeline (llama.cpp for GGUF rows)
Card snapshot main · updated 2026-09-08
malaiwah/Qwen3.8-27B-EXL3-K6-parity
- Reference
- shared BF16 reference capture (vLLM)
- Eval data
- shard 0 of v5 held-out suite: 512 contexts, 1,048,064 scored positions, 330 source clusters
- Context
- not stated
- Tool
- tools/fidelity.py replay; cluster bootstrap 10,000 resamples; cross-engine for GGUF (llama.cpp)
Card snapshot main · updated 2026-08-20
malaiwah/Qwen3.8-27B-EXL3-S16-V-research
- Reference
- shared BF16 reference capture, shared BF16 head
- Eval data
- shard 0 of v5 held-out suite, 512 contexts, 1,048,064 scored positions, 330 source clusters
- Context
- not stated
- Tool
- hidden state captured after final norm, replayed through BF16 lm_head; token_mean_kld, cluster bootstrap 10,000 resamples
Card snapshot main · updated 2026-08-20
malekoo/Qwen3.8-27B-MLX-4bit
- Reference
- bf16 MLX conversion in identical harness
- Eval data
- wikitext-2 (test)
- Context
- disjoint 2048-token windows, single token stream
- Tool
- mlx_lm
Card snapshot main · updated 2026-08-14
malekoo/Qwen3.8-27B-MLX-8bit
- Reference
- bf16 MLX conversion in the identical harness
- Eval data
- wikitext-2 test
- Context
- disjoint 2048-token windows, single token stream
- Tool
- mlx-lm (mlx_lm.convert), M5 Max MacBook Pro
Card snapshot main · updated 2026-08-14
malekoo/Qwen3.8-27B-MLX-bf16
- Reference
- none (perplexity only)
- Eval data
- wikitext-2 test
- Context
- disjoint 2048-token windows, single token stream
- Tool
- mlx-lm harness on Apple silicon (Metal)
Card snapshot main · updated 2026-08-14
manitcor/Qwen3.8-27B-Obliterated-E03-bnb-4bit
- Reference
- BF16 E03 parent checkpoint
- Eval data
- untouched 200-prompt final-test partition; 33 prompts for first-token KL; small capability suite
- Context
- not stated
- Tool
- BitsAndBytes NF4 (0.50.0), compute dtype BF16
Card snapshot main · updated 2026-09-03
mauromedda/fucina-qwen38-v2
- Reference
- not stated
- Eval data
- tool-eval-bench 2.0.4, 69-case
- Context
- not stated
- Tool
- tool-eval-bench 2.0.4 on stock SGLang v0.5.20
Card snapshot main · updated 2026-09-24
MESHIVEAI/Qwen-3.8-27B-Q4_K_M-Imatrix
- Reference
- F16 GGUF baseline
- Eval data
- Wikitext-2 test (296,960 tokens, 580 chunks); Korean Wikipedia 20231101.ko (1,123,328 tokens) perplexity only
- Context
- 512 (-c 512)
- Tool
- llama-perplexity with --kl-divergence
Card snapshot main · updated 2026-09-24
Mia-AiLab/Qwen3.8-27B-EXL3-3.5bpw
- Reference
- none (absolute agentic benchmark)
- Eval data
- tool-eval-bench 2.5.1 hardmode, 84 scenarios; passkey/needle gates
- Context
- not stated
- Tool
- tool-eval-bench 2.5.1, temperature 1.0, seed 42
Card snapshot main · updated 2026-08-25
MiawTeam/Qwen3.8-27B-Uncensored-GGUF
- Reference
- unmodified Qwen3.8-27B base (for capability and refusal deltas); no BF16 reference for quant PPL
- Eval data
- wikitext-2 (PPL); MMLU, ARC-Challenge, HellaSwag, Winogrande 0-shot; 100 held-out harmful prompts
- Context
- not stated
- Tool
- lm-evaluation-harness (bf16, 0-shot); Heretic abliteration search; llama.cpp perplexity
Card snapshot main · updated 2026-08-15
Mihai-LeanZero/Qwen3.8-27B-Atlassian-Q6_K-GGUF
- Reference
- bf16 GGUF's own logits (merged Atlassian adapter, bf16)
- Eval data
- 40 chunks of held-out Atlassian text
- Context
- 2048 tokens per chunk
- Tool
- llama-perplexity --kl-divergence
Card snapshot main · updated 2026-09-17
Mihai-LeanZero/Qwen3.8-27B-Atlassian-Q8_0-GGUF
- Reference
- bf16 GGUF's own logits (merged Atlassian adapter)
- Eval data
- held-out Atlassian text, 40 chunks of 2048 tokens
- Context
- 2048
- Tool
- llama-perplexity --kl-divergence
Card snapshot main · updated 2026-09-17
Minachist/Qwen3.8-27B-INT5-Flat-5.7bpw-AutoRound
- Reference
- BF16 base, KL(P_base || Q_quant) in nats
- Eval data
- WikiText-2 test, 240 x 512 tokens, 122,640 scored positions
- Context
- 512
- Tool
- AutoRound, calibration pile-10k + github-code-clean
Card snapshot main · updated 2026-08-26
Minachist/Qwen3.8-27B-INT6-Flat-6.6bpw-AutoRound
- Reference
- BF16 base model
- Eval data
- WikiText-2 test, 240 sequences x 512 tokens, 122,640 scored positions
- Context
- 512
- Tool
- not stated
Card snapshot main · updated 2026-08-26
Minachist/Qwen3.8-27B-INT6-Mixed-AutoRound
- Reference
- BF16 base model
- Eval data
- WikiText-2 test, 240 sequences x 512 tokens, 122,640 scored positions
- Context
- 512 tokens
- Tool
- KL(P_base || Q_quant) in nats (custom)
Card snapshot main · updated 2026-08-21
Minachist/Qwen3.8-27B-INT8-AutoRound
- Reference
- Qwen/Qwen3.8-27B BF16 teacher, KL over full 248,320 vocab, FP32 log_softmax
- Eval data
- WikiText-2 raw test, 240 seq x 512 = 122,640 scored positions
- Context
- 512
- Tool
- AutoRound (calibration pile-10k + github-code-clean)
Card snapshot main · updated 2026-08-20
minima-ai/mnma_qwen3.8_27b_nvfp4
- Reference
- BF16 model served identically
- Eval data
- WikiText-2 PPL @4K/@32K, MMLU-Pro, GSM8K, AIME'25, GPQA-Diamond, LiveCodeBench v6, RULER NIAH
- Context
- 4K and 32K for perplexity
- Tool
- vLLM 0.27.1, TP=1, one 96 GB RTX PRO 6000, FP8 KV cache
Card snapshot main · updated 2026-09-04
mlasli/Qwen3.8-27B-Heretic-Abliterated-MLX-6bit
- Reference
- base Qwen/Qwen3.8-27B, bf16, greedy
- Eval data
- MMLU, ARC-Challenge, HellaSwag, Winogrande, GSM8K
- Context
- not stated
- Tool
- not stated (few-shot harness)
Card snapshot main · updated 2026-09-01
mlx-bun/Qwen3.8-27B-Trellis-3.2bpw
- Reference
- BF16 base figures from Red Hat's published evaluation (three-seed averages), not measured by author
- Eval data
- GSM8K-Platinum, IFEval, GPQA Diamond; full datasets, zero-shot, thinking enabled
- Context
- seed 1234, MTP off, one seed
- Tool
- mlx-bun (refactor/monorepo at 7c6155b7)
Card snapshot main · updated 2026-10-11
mlx-community/Qwen3.8-27B-OptiQ-4bit
- Reference
- not stated (no BF16 comparison column)
- Eval data
- MMLU 5-shot (1000), GSM8K 3-shot CoT, IFEval, BFCL-V3 simple, HumanEval, HashHop
- Context
- not stated
- Tool
- mlx-optiq
Card snapshot main · updated 2026-09-16
mtrskeen/qwen3.8-27b-ninfer-minq4-mtpq4-visionq4-5080
- Reference
- simulated PPL gate against the IQ3_XXS anchor (no BF16 reference stated)
- Eval data
- gsm8k_cot 8-shot 50 samples; RULER NIAH 8K/32K/64K
- Context
- not stated
- Tool
- lm-evaluation-harness; ninfer-5080 fork
Card snapshot main · updated 2026-09-18
mudler/Qwen3.8-27B-APEX-GGUF
- Reference
- BF16 reference logits
- Eval data
- wikitext-2-raw, 200 chunks
- Context
- n_ctx 512
- Tool
- not stated (imatrix from diverse calibration data, 200 chunks; 19 arms)
Card snapshot main · updated 2026-08-17
MuXodious/Qwen3.8-27B-absolute-heresy.rc12-GGUF
- Reference
- KL divergence reference not stated
- Eval data
- piqa, copa, winogrande 0-shot
- Context
- not stated
- Tool
- lm-eval at Q4_K_M
Card snapshot main · updated 2026-08-15
nerkyor/Qwen3.8-27B-Uncensored-Heretic-T10-BF16
- Reference
- none (absolute scores of the abliterated model; no base-model scores on card)
- Eval data
- MMLU500, GPQA198, LCB100, Refusal140
- Context
- not stated
- Tool
- llama.cpp Q8_0 frozen results
Card snapshot main · updated 2026-08-20
neroued/Qwen3.8-27B-nvfp4-NInfer
- Reference
- official Qwen3.8-27B BF16 figures from the upstream model card (not same protocol)
- Eval data
- IFBench, AIME 2025, AIME 2026, GPQA-Diamond, ERQA, RealWorldQA
- Context
- 252,928-token limit text suite; 81,920 multimodal
- Tool
- NInfer EvalScope 1.9.0, 0-shot rule scoring, thinking on, MTP=3, INT8 group-64 KV, temp 1.0 top-p 0.95 top-k 20, seed 42
Card snapshot main · updated 2026-09-29
nibor1896/Qwen3.8-27B-CNQ4.5
- Reference
- upstream BF16 Qwen3.8-27B weights through crow-nest's f32 oracle
- Eval data
- 3,470 rows = 905 code + 2,565 prose
- Context
- not stated (short rows)
- Tool
- crow-nest tools/oracle-kld.py, KL(reference || arm) in nats, mean +- SE; RTX 5090
Card snapshot main · updated 2026-09-27
nicosuter/Qwen3.8-27B-AWQ
- Reference
- FP8 checkpoint of Qwen3.8-27B (same items, same order); recovery = AWQ/FP8 geometric mean
- Eval data
- BFCL, GPQA Diamond, LiveCodeBench v6, MMMU-Pro, Multimodal (DocVQA/ChartQA/TextVQA), RULER (not counted)
- Context
- RULER synthesized at 4k, 32k, 128k
- Tool
- EVAL.md protocol on H200 NVL; bootstrap 20,000 resamples
Card snapshot main · updated 2026-08-24
nm-testing/Qwen3.8-27B-MXFP6A8
- Reference
- Qwen/Qwen3.8-27B (BF16)
- Eval data
- GSM8K, 5-shot, all 1319 questions
- Context
- not stated
- Tool
- vLLM tests/evals/gsm8k/gsm8k_eval.py on one B300, greedy
Card snapshot main · updated 2026-10-14
nom666/Qwopus3.8-27B-MTPLX-8bit-Quality
- Reference
- Qwen3.8-27B 8-bit, same harness
- Eval data
- 13-task hard suite (thinking); non-thinking suite
- Context
- not stated
- Tool
- MTPLX on Apple M5 Max
Card snapshot main · updated 2026-08-20
notnaurw/Qwen3.8-27B-oQ5e-SAGE-ARCS-mtp
- Reference
- bf16 source model, shared bf16 reference cache
- Eval data
- agentic basis: 96 seq x 2048 tokens, 48 batches, 196,512 positions; stock basis: 128 seq x 1024 tokens, 32 batches, 130,944 positions; long-context 2K/8K/16K
- Context
- 2048 (agentic), 1024 (stock), 2K/8K/16K (long context)
- Tool
- head-KL = KL(bf16 || quantized) over bf16 top-1024 logits, renormalized; full-vocabulary KL as audit
Card snapshot main · updated 2026-10-11
notnaurw/Qwen3.8-27B-oQ5e-SAGE-CREST-mtp
- Reference
- bf16 source model, KL(bf16 || quantized) over bf16 top-1024 logits (head-KL), full-vocab KL as audit
- Eval data
- agentic SWE/terminal corpus (96 seq x 2048 tok); generalist/stock corpus (128 seq x 1024 tok); 16 held-out 131072-token sessions
- Context
- 2048 / 1024 / 131072
- Tool
- not stated
Card snapshot main · updated 2026-10-11
notnaurw/Qwen3.8-27B-oQ5e-SAGE-mtp
- Reference
- bf16 source checkpoint (common reference cache)
- Eval data
- 32 held-out batches / 130,944 tokens; full-vocab audit 64 x 1024 tokens
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-10-11
notnaurw/Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp
- Reference
- bf16 source (Qwen3.8-27B), shared bf16 reference cache
- Eval data
- agentic SWE/terminal corpus (96 seq x 2048 tok, 48 batches); generalist/stock corpus (128 seq x 1024 tok, 32 batches); 16 held-out agent sessions at 131072 tok
- Context
- 2048 / 1024 / 131072 tokens
- Tool
- head-KL over bf16 top-1024 logits plus full-vocabulary KL audit (custom harness)
Card snapshot main · updated 2026-10-12
nova-quill/OBLITERATUS-Qwen3.8-27B-OBLITERATED-unsloth-mlx
- Reference
- stock Qwen3.8-27B
- Eval data
- MMLU, n=50 per subject (2,850 questions); full 14k run in progress
- Context
- 0-shot
- Tool
- lm-eval-harness
Card snapshot main · updated 2026-08-24
okayuji/Qwen3.8-27B-coreml-agent
- Reference
- bf16 reference implementation
- Eval data
- 10 prompts x 64 positions (author's lab record, not published)
- Context
- 16,384-token context build
- Tool
- Core ML runtime / agent benchmark
Card snapshot main · updated 2026-10-05
Omega-Factory/Qwen3.8-27B-GGUF
- Reference
- FP16 base Qwen/Qwen3.8-27B @ 1d4bf0f2
- Eval data
- holdout y_exam, N=200 per category (ARC-MC, GSM8K, MATH-500)
- Context
- not stated
- Tool
- deterministic eval with McNemar item-level flips
Card snapshot main · updated 2026-10-11
oracomputing/Qwen3.8-27B-OQ3-GGUF
- Reference
- original 16-bit Qwen/Qwen3.8-27B
- Eval data
- MMLU-Pro (5-shot CoT generative, 12,032 questions); AIME-25 (8 seeds x 30 problems, 65,536 max new tokens)
- Context
- not stated
- Tool
- not stated; same setup for all models, matched total byte budget (10.17 vs 10.15 GB live weights)
Card snapshot main · updated 2026-08-26
orcarouter/OrcaSAQ-2-27B
- Reference
- BF16 Qwen3.8-27B
- Eval data
- WikiText-2, 16,376 predicted tokens
- Context
- not stated
- Tool
- not stated (same evaluation path as BF16)
Card snapshot main · updated 2026-10-12
orcarouter/Qwen3.8-27B-MLX
- Reference
- BF16
- Eval data
- wikitext-2
- Context
- 1024 tokens
- Tool
- MLX CUDA backend
Card snapshot main · updated 2026-10-12
OS-Software/Hemmingway-1-Uncensored-Heretic-GGUF
- Reference
- Altworld/Hemmingway-1 (original model)
- Eval data
- not stated
- Context
- not stated
- Tool
- Heretic v2.0.0.dev0+custom
Card snapshot main · updated 2026-09-24
OS-Software/Qwen3.8-27B-MTP-heretic-ja-GGUF
- Reference
- unsloth/Qwen3.8-27B (original model)
- Eval data
- not stated
- Context
- not stated
- Tool
- Heretic v1.4.0+custom, ARA method
Card snapshot main · updated 2026-08-16
OS-Software/Qwen3.8-27B-Uncensored-Heretic-v2-GGUF
- Reference
- Original Qwen/Qwen3.8-27B
- Eval data
- new custom mixed-language dataset (Japanese-language sets used for KLD and refusal rate)
- Context
- not stated
- Tool
- Heretic v1.4.0+custom, ARA method
Card snapshot main · updated 2026-08-31
OS-Software/Qwen3.8-27B-Uncensored-Heretic-v3-GGUF
- Reference
- Qwen/Qwen3.8-27B (original model)
- Eval data
- not stated
- Context
- not stated
- Tool
- Heretic v2.0.0.dev0+custom
Card snapshot main · updated 2026-09-11
oshiteku/Sanshoku-Qwen3.8-27B-GGUF
- Reference
- BF16
- Eval data
- 60 in-house prompts, 32 tokens each (top-1/KL); fixed subsets of GSM8K (100), MATH-500 (50), MGSM Japanese (100) for Math; SWE-bench Verified Mini (50 tasks) with mini-SWE-agent
- Context
- not stated
- Tool
- not stated for top-1/KL; llama.cpp 1a67982 and PrismML fork bdc23b5 tested
Card snapshot main · updated 2026-09-23
pearsonkyle/Qwen3.8-27B-GPTQ-W4A16
- Reference
- bf16 served identically on vLLM
- Eval data
- 174 held-out tool-calling turns; KLD over 6 distributions (agentic text lowest, general English highest)
- Context
- not stated for eval (calibration at ctx 32,768)
- Tool
- vLLM
Card snapshot main · updated 2026-09-03
pearsonkyle/Qwen3.8-27B-imatrix-MTP-GGUF
- Reference
- FP16
- Eval data
- external holdout (code + math + tools, 90k tokens, disjoint from calibration) plus five other eval distributions each with own FP16 baseline
- Context
- eval-ctx 8192
- Tool
- llama-perplexity (Quant-Tuner, llama.cpp f3e1828)
Card snapshot main · updated 2026-08-15
penkia/TernaryQuench-Qwen3.8-27B-GGUF
- Reference
- Qwen3.8-27B BF16 (five-task); Qwen3.8-27B 4-bit MLX (PPL)
- Eval data
- WikiText-2 test, first 20,480 tokens as 40 independent 512-token chunks; PIQA, ARC-Easy, ARC-Challenge, HellaSwag, WinoGrande full sets
- Context
- 512-token chunks
- Tool
- llama.cpp (Q2_K file) for PPL; lm-eval 0.4.7 zero-shot on llama.cpp CUDA FP32 acc; reference rows MLX / native PyTorch
Card snapshot main · updated 2026-09-07
penkia/TernaryQuench-Qwen3.8-27B-MLX
- Reference
- not stated (absolute perplexity and zero-shot accuracy, no BF16 delta)
- Eval data
- WikiText-2 test sample (first 20,480 tokens, 40 chunks); lm-eval PIQA/ARC/HellaSwag/WinoGrande
- Context
- 512-token chunks
- Tool
- MLX; lm-eval 0.4.7 zero-shot
Card snapshot main · updated 2026-09-07
perkel/Qwen3.8-27B-MC
- Reference
- Qwen BF16 weights computed with FP32 math
- Eval data
- 81,880 held-out tokens: reasoning traces, math, code, chat, wiki (disjoint from calibration)
- Context
- not stated (engine at 16-bit activations, unquantized KV cache)
- Tool
- MegaCapybara engine; GPTQ with MX block scales, 514K calibration tokens
Card snapshot main · updated 2026-10-07
philbert440/Qwen3.8-27B-W4A16-AWQ
- Reference
- not stated (no bf16 comparison of the quant; base-model benchmarks quoted from Qwen card, not re-measured)
- Eval data
- Qwen model-card benchmarks (quoted); GSM8K sample 3/3 sanity check
- Context
- not stated
- Tool
- not stated (llm-compressor AWQ calibration; vLLM speed benchmarks)
Card snapshot main · updated 2026-08-15
Pilcothink/Qwen3.8-27B-MixedInt2-ultra-lite
- Reference
- Qwen/Qwen3.8-27B
- Eval data
- MMLU acc, GSM8K exact_match (flexible and strict)
- Context
- not stated
- Tool
- AutoRound evaluation interface with LM Evaluation Harness
Card snapshot main · updated 2026-08-24
Pilcothink/Qwen3.8-27B-MixedInt4-AutoRound
- Reference
- original Qwen3.8-27B
- Eval data
- MMLU, GSM8K, ARC-Challenge, BoolQ, HellaSwag, PIQA, WinoGrande
- Context
- not stated
- Tool
- AutoRound evaluation interface with LM Evaluation Harness
Card snapshot main · updated 2026-08-17
PocketAiHub/Qwen3.8-27B-Abliterated-MTPLX-Optimized-Speed
- Reference
- regular (non-abliterated) MTPLX at matched quantization layout; BF16-greedy teacher-forced positions
- Eval data
- 24 prompts (12 capability + 12 JailbreakBench), 16 positions each, 384 positions
- Context
- not stated
- Tool
- full 248,320-vocab logits, float32 capture, float64 math, forward KL in nats
Card snapshot main · updated 2026-08-22
poisonxa/Qwen3.8-27B-PXQN-OneCard
- Reference
- Q8_0 file of the same model
- Eval data
- held-out chat-format corpus, assistant-role tokens only (15,065 tokens) for KLD; wikitext-2 test set for PPL
- Context
- KLD: 24 chunks of 2048 tokens; PPL: 200 chunks of 512 tokens
- Tool
- PXA v2026.10 engine
Card snapshot main · updated 2026-10-12
PollardWeights/Penjing-27B-Pollard
- Reference
- f16 (PPL 4.1422)
- Eval data
- wikitext2 test
- Context
- -c 2048
- Tool
- llama-perplexity (Same top p for top-1 agree)
Card snapshot main · updated 2026-09-22
primitive-ai/Qwen3.8-27B-mixed-NVFP4-FP8
- Reference
- Qwen BF16 (Qwen/Qwen3.8-27B) as accuracy baseline; no KL/PPL
- Eval data
- 1,370 items over 14 public benchmarks: 1,170-item knowledge suite + 200-item tool-calling suite
- Context
- 16,384-token budget, concurrency 32, thinking forced on
- Tool
- temperature 0.6, top_p 0.95, top_k 20, vLLM, one RTX PRO 6000 Blackwell, all rows back to back
Card snapshot main · updated 2026-08-27
ProCreations/bonsai-2-27b-gsq-rco-gguf
- Reference
- original Bonsai PTQ1_0 quant and original Qwen3.8 BF16
- Eval data
- WikiText-2 raw test; ARC-C 1,172; ARC-E 2,376; HellaSwag fixed random 2,000 validation; Winogrande
- Context
- WikiText-2 context 2048
- Tool
- native perplexity tool; likelihood harness without chat template
Card snapshot main · updated 2026-09-19
pugant/Qwen3.8-27B-MTP-Q3_0_ROCMFPX
- Reference
- BF16 historical references (6.6409 en / 11.7156 it)
- Eval data
- wikitext-2-en 150k tokens + Italian technical corpus 51.5k tokens
- Context
- 512
- Tool
- llama-perplexity -c 512 -b 512 -fa on -ngl 999, Vulkan RADV, Strix Halo
Card snapshot main · updated 2026-09-04
QTML/Qwen3.8-27B-Security-Quark-MXFP4
- Reference
- Base Qwen3.8-27B BF16
- Eval data
- 26-prompt cybersecurity refusal benchmark; benign prompts for KL
- Context
- temperature 0, thinking disabled
- Tool
- Served first-token top-20 truncated logprobs for abliteration-vs-quant table; heretic full-vocab first-token KL for 0.0151; AMD Quark AWQ-MXFP4
Card snapshot main · updated 2026-09-04
quark75/Qwen3.8-27B-EXL3-3.0bpw
- Reference
- original BF16 weights
- Eval data
- held-out evaluation set (not named)
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-09-19
quark75/Qwen3.8-27B-Q4_0.gguf
- Reference
- original BF16 weights
- Eval data
- not stated
- Context
- n_ctx=512, 580 chunks
- Tool
- llama.cpp llama-perplexity --kl-divergence
Card snapshot main · updated 2026-09-19
QUASAR-QAT/Qwen3.8-27B-QUASAR-NVFP4
- Reference
- BF16 original compared on same benchmarks; QAT distillation vs frozen BF16 teacher
- Eval data
- GPQA-Diamond (2 runs, n=396), AIME'26 (3 repeats, n=90), MRCR, HMMT'26, SuperGPQA, LiveCodeBench, RULER
- Context
- not stated
- Tool
- QUASAR (loss-aware NVFP4 QAD); comparators via LLM Compressor
Card snapshot main · updated 2026-10-15
r0b0tlab/Qwen3.8-27B-EXL3-4.00bpw
- Reference
- not stated
- Eval data
- Q200v2 text-180 (gsm8k, humaneval, ifeval); GSM8K greedy for acceptance length
- Context
- not stated
- Tool
- ExLlamaV3 v1.5.0 convert.py
Card snapshot main · updated 2026-09-17
Rabbit-Hole-Ai/Qwen3.8-27B-5090-goldilocks-GGUF
- Reference
- true BF16 reference GGUF (PPL 8.1090)
- Eval data
- 143 chunks of a mixed corpus; also corpus A (literary prose) and GPQA-style science prose
- Context
- n_ctx 512
- Tool
- llama-perplexity --kl-divergence, llama.cpp build 10437
Card snapshot main · updated 2026-09-03
RadixArk/Qwen3.8-27B-NVFP4-BF16-LMHead
- Reference
- not stated (absolute GSM8K accuracy)
- Eval data
- GSM8K full 1,319 split, thinking mode
- Context
- not stated
- Tool
- SGLang main d287880a, TP4 on 4x GB300, temperature 1.0, seed 0
Card snapshot main · updated 2026-08-24
rapid-mlx/Qwen3.8-27B-mixed-3.5bpw-MLX
- Reference
- none for main table; bf16 included in subset comparison
- Eval data
- MMLU-Pro n=500, GSM8K n=150, HumanEval+ n=40, MMBench-EN n=300; identical subset (200 questions / 60 problems / 150 images) for bf16 comparison
- Context
- not stated
- Tool
- identical harness, greedy decoding
Card snapshot main · updated 2026-10-13
rdtand/Qwen3.8-27B-PrismaAQUA-5.5bit-vllm
- Reference
- BF16 source (teacher), top-1024 logprobs plus a lumped residual bucket
- Eval data
- WikiText test (calibration windows from WikiText train), 8 x 512 = 4088 positions
- Context
- 512 tokens
- Tool
- not stated by name (pipeline tool with shipcard.json)
Card snapshot main · updated 2026-08-23
rdtand/Qwen3.8-27B-PrismaAQUA-gridbook-13GB-5080-vllm
- Reference
- unquantized BF16 Qwen3.8-27B text checkpoint
- Eval data
- WikiText-2 (8176 tokens scored for PPL); KL over 4088 positions
- Context
- seqlen 512 for PPL
- Tool
- vLLM served artifact, top-K prompt logprobs via /v1/completions
Card snapshot main · updated 2026-08-18
rdtand/Qwen3.8-27B-PrismaScout-AQUA-Vision-20GB
- Reference
- pinned BF16 Qwen3.8 source
- Eval data
- KL: eight 512-token samples (4,088 positions); PPL: WikiText-2 test, 8,176 tokens in 16 non-overlapping 512-token chunks
- Context
- 512
- Tool
- not stated
Card snapshot main · updated 2026-08-24
RedHatAI/Qwen3.8-27B-MXFP4
- Reference
- not stated (absolute benchmark scores; v2 compares two checkpoints, unnamed in digest)
- Eval data
- GSM8K Platinum, IFEval, MATH-500, AIME 2025, GPQA Diamond (v1); AIME 2026, Terminal-Bench 2.1, MMLU-Pro, DeepSWE (v2)
- Context
- max_length 69632 (lm-eval)
- Tool
- lm-evaluation-harness and lighteval via vLLM on B200 (v1); Inspect-AI on MI355X (v2)
Card snapshot main · updated 2026-10-08
Reza2kn/Bev
- Reference
- gold labels of Jev Persian Benchmark (accuracy vs gold, not a quant-vs-reference comparison)
- Eval data
- Jev Persian Benchmark, 480 main questions, commit ac218d96
- Context
- 16,384 tokens per field (serving config)
- Tool
- Bev v0.1.1 scoring extension on Prism Ternary-Bonsai-2-27B GGUF
Card snapshot main · updated 2026-09-23
RonnieOps/Qwen3.8-27B-uncensored-IQ4_XS
- Reference
- saved BF16 logits (same engine); PPL anchors BF16
- Eval data
- wiki validation corpus (PPL, KL); custom 59-task SEC/finance domain harness; 115/127-task suites
- Context
- 32K for task suites; not stated for PPL/KL
- Tool
- llama.cpp (KL RMS dp / top-p agreement vs BF16)
Card snapshot main · updated 2026-09-21
RukaRat/Qwen3.8-27B-INT8-W8A8-imatrix-heretic-MTP
- Reference
- ground truth answers of an in-house battery (160 items); unmodified W8A8 build as comparison
- Eval data
- in-house battery, four tasks of 40 (incl. NER and extraction); refusal tests on harmful_behaviors and OR-Bench-Hard-1K
- Context
- 286,720 context for serving
- Tool
- vLLM TP=2 on 2x RTX 3090, MTP on, fp8 KV
Card snapshot main · updated 2026-09-19
RukaRat/Qwen3.8-27B-W4A16-imatrix-heretic-int8head-MTP
- Reference
- pairwise vs the publisher's other builds (ground-truth scored 160-item battery); no BF16 reference
- Eval data
- five-task battery of 160 items (200-item version earlier), refusal eval with OR-Bench-Hard-1K, 2 vision probes
- Context
- not stated
- Tool
- vLLM on actual served checkpoint; abliteration with Heretic 1.4.0 before W4A16 quantization
Card snapshot main · updated 2026-09-19
sakamakismile/Qwen3.8-27B-EXL3-6.5bpw
- Reference
- none (absolute scores; castings compared)
- Eval data
- MMLU-Pro first 20 questions; 8-task agent battery x3
- Context
- not stated
- Tool
- ExLlamaV3 v1.4.5, TP=8 on 8 x RTX PRO 2000
Card snapshot main · updated 2026-09-05
sammyblues/Qwen3.8-27B-IQ2_XS-GGUF
- Reference
- 64L BF16 reference (MTP head removed, same as quant)
- Eval data
- WikiText-2 test, all 453 512-token chunks; HellaSwag fixed 100-task subset; imatrix from 128 WikiText-2 train chunks of 512 tokens
- Context
- 512-token PPL chunks; 4K context for memory test
- Tool
- llama.cpp (llama-quantize, perplexity) on A100 80GB PCIe
Card snapshot main · updated 2026-09-02
sdkyuan/qwen3.8-27B-qat-q2_0-gguf
- Reference
- FP16 teacher
- Eval data
- held-out chat-formatted slices (reasoning, code, tool calling) and wikitext-2 test
- Context
- not stated
- Tool
- top-1 agreement and mean KL divergence (nats/token)
Card snapshot main · updated 2026-08-22
shisa-ai/Qwen3.8-27B-FP8-BLOCK
- Reference
- Qwen/Qwen3.8-27B BF16 base @ 1d4bf0f2
- Eval data
- 512 contexts / 1,048,064 positions, full 248,320 vocab, one shared BF16 LM head
- Context
- not stated (shard-0 KLD replay)
- Tool
- KL(BF16 base || candidate), v5 shard-0 KLD replay; all checkpoints captured/replayed locally
Card snapshot main · updated 2026-08-17
sigmanih/Qwen-Qwen3.8-27B-GGUF-IQ1_S
- Reference
- not stated (absolute benchmark accuracy, no reference model)
- Eval data
- slice of ARC-Challenge, BBH, GPQA, GSM8K, HellaSwag, HumanEval, MATH, MBPP, MMLU, MMLU-Pro, TruthfulQA; 100 questions total
- Context
- not stated
- Tool
- Sigma Studio Training Lab, temp 0.0, seed 42
Card snapshot main · updated 2026-10-13
sigmanih/Qwen-Qwen3.8-27B-GGUF-IQ2_S
- Reference
- none (absolute benchmark accuracy, no BF16 comparison)
- Eval data
- slice of 12 benchmark suites, 100 questions total (ARC-Challenge, BBH, GPQA, GSM8K, HellaSwag, HumanEval, MATH, MBPP, MMLU, MMLU-Pro, TruthfulQA)
- Context
- not stated
- Tool
- Sigma Studio Training Lab, seed 42, temp 0.0
Card snapshot main · updated 2026-10-13
sigmanih/Qwen-Qwen3.8-27B-GGUF-IQ2_XXS
- Reference
- absolute benchmark accuracy, no bf16 reference
- Eval data
- small slices of ARC-Challenge, BBH, GPQA, GSM8K, HellaSwag, HumanEval, MATH, MBPP, MMLU, MMLU-Pro, TruthfulQA (100 questions total)
- Context
- not stated
- Tool
- Sigma Studio Training Lab (code_execution, continuation_logprob, cot_generation, letter_logprob; temp 0.0, seed 42)
Card snapshot main · updated 2026-10-13
sigmanih/Qwen-Qwen3.8-27B-GGUF-Q4_K_M
- Reference
- not stated (absolute benchmark scores, no BF16 comparison)
- Eval data
- slices of ARC-Challenge, BIG-Bench Hard, GPQA, GSM8K, HellaSwag, HumanEval, MATH, MBPP, MMLU, MMLU-Pro, TruthfulQA (100 questions total)
- Context
- not stated
- Tool
- Sigma Studio Training Lab, temp 0, seed 42
Card snapshot main · updated 2026-10-15
sigmanih/Qwen-Qwen3.8-27B-GGUF-Q5_K_M
- Reference
- none (absolute benchmark accuracy)
- Eval data
- slice of benchmark datasets (100 questions total), not full suites [L100]
- Context
- not stated
- Tool
- Sigma Studio Training Lab, temp 0, seed 42; protocols code_execution, continuation_logprob, cot_generation, letter_logprob
Card snapshot main · updated 2026-10-16
signalnine/Qwen3.8-27B-MTP-q27
- Reference
- not stated (v2 recipes vs 3.6-recipe builds; q6k reference for long-context tail)
- Eval data
- wikitext PPL; HumanEval+ 30 problems; needle-in-haystack
- Context
- not stated
- Tool
- q27 engine
Card snapshot main · updated 2026-09-09
singulared/Qwen3.8-27B-ROCmFP4-MTP-GGUF
- Reference
- mainline llama.cpp Q4_K_M (no BF16 reference)
- Eval data
- held-out wikitext-2, 145 chunks
- Context
- 2048
- Tool
- llama-perplexity
Card snapshot main · updated 2026-08-19
slevinw/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF
- Reference
- RVN-F16.gguf (F16 reference) for perplexity table
- Eval data
- tiny_shakespeare corpus
- Context
- 2048
- Tool
- llama-perplexity on RTX PRO 6000 Blackwell, llama.cpp master
Card snapshot main · updated 2026-08-24
SmokeyPete/Qwen3.8-27B-Heretic-NVFP4-DGX-Spark
- Reference
- BF16 = the abliterated model before quantization
- Eval data
- GSM8K (1,319 questions), HumanEval+ (164, EvalPlus 0.3.1)
- Context
- not stated
- Tool
- SGLang 0.5.19, greedy, thinking on reasoning_effort medium, no speculative decoding
Card snapshot main · updated 2026-10-03
soyrsoyr/Qwen3.8-27B-W4A16-AWQ-GPTQ
- Reference
- bf16 base Qwen3.8-27B
- Eval data
- OpenLLM v1 (ARC-Challenge, HellaSwag, TruthfulQA-mc2, Winogrande); reasoning suite AIME-24, AIME-25 avg@4, MATH-500 pass@1
- Context
- generative up to 32k tokens, temp 0.6, top_p 0.95
- Tool
- lm-evaluation-harness and lighteval through vLLM
Card snapshot main · updated 2026-08-18
staccs/Qwen3.8-27B-lecore-GGUF
- Reference
- not stated (weights dequantized to bf16 for evaluation)
- Eval data
- MMLU 5-shot first 10 questions of each of 57 subjects; GSM8K 5-shot first 100 test problems
- Context
- not stated
- Tool
- lm-eval-harness 0.4.9, greedy, 320 new tokens, no chat template
Card snapshot main · updated 2026-10-15
staccs/Qwen3.8-27B-lecore-mlx
- Reference
- bf16 reference
- Eval data
- WikiText-2 test (32704 tok) and a different Python-source slice (8176 tok); MMLU 5-shot first 10 questions of each of 57 subjects; GSM8K 5-shot first 100 test problems
- Context
- seq 512
- Tool
- MLX perplexity on dequantized shipped weights; lm-eval-harness 0.4.9
Card snapshot main · updated 2026-10-15
Steven10429/Qwen3.8-27B-UD-IQ4_XS-MLX
- Reference
- direct bf16 to 4-bit MLX conversion (lmstudio-community)
- Eval data
- wikitext-2
- Context
- 16 windows x 2048 tokens
- Tool
- not stated
Card snapshot main · updated 2026-10-05
stosc/qwen3_8_27b_nvfp4_v100.ninfer
- Reference
- official Qwen3.8-27B BF16 figures from upstream card (not same protocol)
- Eval data
- IFBench, AIME 2025, AIME 2026, GPQA-Diamond, ERQA, RealWorldQA; 1,723 samples, single sample each, not pass@k
- Context
- 252,928-token context cap for text suite; 81,920 for multimodal suite; thinking on, MTP=3, INT8 group-64 KV
- Tool
- NInfer EvalScope 1.9.0, 0-shot, rule-based scoring, temp 1.0, top-p 0.95, top-k 20, seed 42
Card snapshot main · updated 2026-09-26
sunnyyy/Qwen3.8-27B-NInfer
- Reference
- official Qwen3.8-27B BF16 figures from upstream card (not same protocol)
- Eval data
- IFBench, AIME 2025, AIME 2026, GPQA-Diamond, ERQA, RealWorldQA (1,723 samples)
- Context
- 262,144 text limit; 81,920 multimodal
- Tool
- NInfer EvalScope 1.9.0, 0-shot, thinking on, MTP=3, INT8 group-64 KV, temp 1.0 top-p 0.95 top-k 20, seed 42
Card snapshot main · updated 2026-09-01
sunnyyy/Qwen3.8-27B-nvfp4-NInfer
- Reference
- none; compared against upstream official BF16 figures (not same protocol)
- Eval data
- IFBench, AIME 2025, AIME 2026, GPQA-Diamond, ERQA, RealWorldQA (1,723 samples)
- Context
- 252,928-token limit text suite; 81,920 multimodal
- Tool
- EvalScope 1.9.0 via NInfer, thinking on, MTP=3, INT8 group-64 KV, temp 1.0, top-p 0.95, top-k 20, seed 42, one sample per problem
Card snapshot main · updated 2026-09-01
syvai/qwen3.8-27b-3090-fast-variant
- Reference
- bf16
- Eval data
- GPTQ calibrated on 300k of the model's own hidden states
- Context
- not stated
- Tool
- GPTQ int4, group 128
Card snapshot main · updated 2026-08-18
talxcc/Tals-coder-flash-01
- Reference
- unsloth Q4_K_M (100% baseline)
- Eval data
- HumanEval, MBPP, multi-file generation, refactoring; MMLU, GSM8k (named, no per-bench scores)
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-09-29
tcclaviger/Qwen3.8-27B-RFI8-AA
- Reference
- BF16 original
- Eval data
- WikiText-2
- Context
- n_ctx 2048
- Tool
- PPL tool included with the checkpoint
Card snapshot main · updated 2026-08-16
TechnoBaptist/Ternary-Bonsai-2-27B-gguf
- Reference
- Qwen3.8-27B FP16 reference
- Eval data
- 14 benchmarks across six skill categories (MMLU-Redux, MuSR, GSM8K, MATH-500, AIME25, AIME26, HumanEval+, MBPP+, LiveCodeBench, IFEval, IFBench, BFCL v3, MMMU-Pro, OCR Bench v2), thinking mode
- Context
- not stated
- Tool
- EvalScope + vLLM on NVIDIA H100
Card snapshot main · updated 2026-09-17
TelperionAI/Qwen3.8-27B-EXL3-4.0bpw
- Reference
- BF16 base
- Eval data
- 231 documents, 142,727 scored token positions
- Context
- not stated
- Tool
- next-token disagreement vs BF16 base, bucketed by base top1-top2 logprob margin; paired McNemar n=33,812
Card snapshot main · updated 2026-09-02
TelperionAI/Qwen3.8-27B-EXL3-4.0bpw-v2
- Reference
- BF16 base Qwen3.8-27B
- Eval data
- 231 documents, 142,727 scored positions (own corpus); n=33,812 in confident bucket
- Context
- not stated
- Tool
- paired McNemar significance; divmed = median first-divergence token under free greedy decode
Card snapshot main · updated 2026-09-04
TelperionAI/Qwen3.8-27B-EXL3-5.5bpw
- Reference
- BF16 base
- Eval data
- 231 documents, 142,727 scored token positions
- Context
- not stated
- Tool
- next-token disagreement vs BF16 base, bucketed by base top1-top2 logprob margin; paired McNemar n=33,812
Card snapshot main · updated 2026-09-02
TelperionAI/Qwen3.8-27B-EXL3-5.5bpw-v2
- Reference
- BF16 base Qwen3.8-27B
- Eval data
- 231 documents, 142,727 scored positions (own corpus); n=33,812 in confident bucket
- Context
- not stated
- Tool
- paired McNemar significance; divmed = median first-divergence token under free greedy decode
Card snapshot main · updated 2026-09-04
TelperionAI/Qwen3.8-27B-EXL3-6.5bpw
- Reference
- BF16 base
- Eval data
- 231 documents, 142,727 scored token positions
- Context
- not stated
- Tool
- next-token disagreement vs BF16 base, bucketed by base top1-top2 logprob margin; paired McNemar n=33,812
Card snapshot main · updated 2026-09-02
TelperionAI/Qwen3.8-27B-FP8-block-AWQ
- Reference
- BF16 base
- Eval data
- 231 documents, 142,727 scored token positions
- Context
- not stated
- Tool
- disagreement rates vs BF16 base bucketed by base top1-top2 logprob margin; paired McNemar significance
Card snapshot main · updated 2026-09-02
TelperionAI/Qwen3.8-27B-INT4-AWQ-GPTQ
- Reference
- BF16 base
- Eval data
- 142,727 tokens of self-distilled thinking-mode output plus 200 free greedy generations
- Context
- not stated
- Tool
- vLLM 0.27.1, TP=2, 2xB300
Card snapshot main · updated 2026-09-02
TelperionAI/Qwen3.8-27B-INT4-AWQ-GPTQ-gdn4
- Reference
- BF16 base
- Eval data
- 142,727 tokens of self-distilled thinking-mode output plus 200 free greedy generations
- Context
- not stated
- Tool
- vLLM 0.27.1, TP=2, 2xB300; perplexity deliberately excluded
Card snapshot main · updated 2026-09-03
TelperionAI/Qwen3.8-27B-NVFP4-AWQ-AutoRound
- Reference
- BF16 base
- Eval data
- 142,727 tokens of self-distilled thinking-mode output plus 200 free greedy generations
- Context
- not stated
- Tool
- vLLM 0.27.1, TP=2, 2xB300 (custom top-1/bucket disagreement metrics; perplexity deliberately excluded)
Card snapshot main · updated 2026-09-02
TelperionAI/Qwen3.8-27B-NVFP4-AWQ-AutoRound-allfp4
- Reference
- BF16 base
- Eval data
- 142,727 tokens of self-distilled thinking-mode output plus 200 free greedy generations
- Context
- not stated
- Tool
- vLLM 0.27.1, TP=2, 2xB300 (custom top-1/bucket disagreement metrics; perplexity deliberately excluded)
Card snapshot main · updated 2026-09-03
TelperionAI/Qwen3.8-27B-NVFP4-AWQ-GPTQ
- Reference
- BF16 base model
- Eval data
- 142,727 tokens of self-distilled thinking-mode output plus 200 free greedy generations
- Context
- not stated
- Tool
- vLLM 0.27.1, TP=2, 2xB300
Card snapshot main · updated 2026-09-03
ThakiCloud/Qwen3.8-27B-NVFP4-FP8ATTN
- Reference
- Qwen/Qwen3.8-27B bf16, paired comparisons with McNemar / p-values
- Eval data
- MMMU-val MC (232 paired), GSM8K n=600, HAERAE, KoBEST, belebele-ko, ARC-Challenge, IFEval
- Context
- not stated
- Tool
- llm-compressor GPTQ oneshot, 1,024 text calibration samples; vLLM 0.27.1 + lm-eval 0.4.12
Card snapshot main · updated 2026-09-07
ThakiCloud/Qwen3.8-27B-NVFP4-GPTQ-mm
- Reference
- Qwen/Qwen3.8-27B bf16
- Eval data
- MMMU validation multiple-choice (246 paired items); GSM8K full 1,319 items on the -txt sibling only
- Context
- max_new_tokens=16384, temperature 0 (MMMU)
- Tool
- llm-compressor GPTQ oneshot NVFP4 on 1,024 text+image calibration samples; vLLM, single B200; lm-eval 0.4.12 for GSM8K
Card snapshot main · updated 2026-09-07
ThakiCloud/Qwen3.8-27B-NVFP4-GPTQ-txt
- Reference
- Qwen/Qwen3.8-27B (bf16)
- Eval data
- MMMU-val multiple-choice (246 paired items); GSM8K full 1,319; HAERAE, KoBEST, belebele-ko, ARC-Challenge (n=600 per task); IFEval prompt-strict
- Context
- max_model_len=32768; MMMU max_new_tokens=16384
- Tool
- lm-eval 0.4.12 with vLLM v0.27.1 on a single B200
Card snapshot main · updated 2026-09-07
ThakiCloud/Qwen3.8-27B-W4A16-AWQ
- Reference
- bf16 base and publisher's GPTQ sibling
- Eval data
- GSM8K full 1,319 items, English and Korean
- Context
- not stated
- Tool
- greedy, one serving session, identical config across arms; strict-match and flexible-extract scoring
Card snapshot main · updated 2026-09-07
ThakiCloud/Qwen3.8-27B-W4A16-GPTQ
- Reference
- bf16 (same serving session)
- Eval data
- GSM8K full 1,319 items en/ko
- Context
- greedy
- Tool
- vLLM 0.27.1 + lm-eval 0.4.12; GPTQ W4A16 via llmcompressor, calibration 1,024 samples x 2,048 tokens
Card snapshot main · updated 2026-09-07
thederby/Qwen3.8-27B-oQ4e-fp16-mtp
- Reference
- original bf16 model
- Eval data
- 48 windows of English prose, Python code, Japanese and Korean text
- Context
- 2048 tokens per window
- Tool
- mean KL on reference top-256 tokens plus remainder bucket (slightly underestimates full KL)
Card snapshot main · updated 2026-10-11
thederby/Qwen3.8-27B-oQ4e-mtp
- Reference
- original bf16 model; mean KL on reference top-256 tokens plus one remainder bucket (slightly underestimates full KL)
- Eval data
- 48 windows of English prose, Python code, Japanese and Korean text
- Context
- 2048 tokens per window
- Tool
- not stated
Card snapshot main · updated 2026-10-11
thederby/Qwen3.8-27B-oQ5e-fp16-mtp
- Reference
- original bf16 model
- Eval data
- 48 windows of English prose, Python code, Japanese and Korean text
- Context
- 2048 tokens per window
- Tool
- mean KL on reference top-256 tokens plus one remainder bucket (slightly underestimates full KL); fidelity not task benchmarks
Card snapshot main · updated 2026-10-11
thederby/Qwen3.8-27B-oQ5e-mtp
- Reference
- original bf16 model
- Eval data
- English prose, Python code, Japanese and Korean text (48 windows)
- Context
- 2048 tokens per window
- Tool
- mean KL on reference top-256 tokens plus one remainder bucket (oMLX-based; tool not named)
Card snapshot main · updated 2026-10-11
thederby/Qwen3.8-27B-oQ6e-fp16-mtp
- Reference
- original bf16 model
- Eval data
- 48 windows x 2048 tokens of English prose, Python code, Japanese and Korean text
- Context
- 2048-token windows
- Tool
- mean KL over reference top-256 tokens plus remainder bucket (slightly underestimates full KL); fidelity not task benchmarks
Card snapshot main · updated 2026-10-11
thederby/Qwen3.8-27B-oQ6e-mtp
- Reference
- original bf16 model
- Eval data
- 48 windows of English prose, Python code, Japanese and Korean text; <0.3% appears in calibration data
- Context
- 2048 tokens per window
- Tool
- mean KL on reference top-256 tokens plus one remainder bucket (slightly underestimates full KL); oMLX oQe
Card snapshot main · updated 2026-10-11
thederby/Qwen3.8-27B-oQ8e-fp16-mtp
- Reference
- original bf16 model
- Eval data
- 48 windows x 2048 tokens of English prose, Python code, Japanese and Korean text
- Context
- 2048-token windows
- Tool
- mean KL on reference top-256 tokens plus remainder bucket (slightly underestimates full KL)
Card snapshot main · updated 2026-10-11
thederby/Qwen3.8-27B-oQ8e-mtp
- Reference
- original bf16 model
- Eval data
- 48 windows x 2048 tokens of English prose, Python code, Japanese and Korean text
- Context
- 2048-token windows
- Tool
- mean KL over reference top-256 tokens plus remainder bucket (slightly underestimates full KL); oMLX oQe
Card snapshot main · updated 2026-10-11
TheDrainFlorist/Qwen3.8-27B-VQ-3.9bpw
- Reference
- bf16 teacher (cached top-64)
- Eval data
- three house corpora: prose, public code, literary (old table: prose only)
- Context
- 12288 tokens (old table: 2048 tokens)
- Tool
- unmodified mlx-lm / KL instrument; tool name not otherwise stated
Card snapshot main · updated 2026-10-11
TheDrainFlorist/Qwen3.8-27B-VQ-4.5bpw
- Reference
- bf16 teacher (cached top-64)
- Eval data
- three corpora: prose, public code, literary; older table prose-only
- Context
- 12288 tokens (current table); 2048 tokens (superseded table)
- Tool
- KL instrument via mlx-lm; millinats per token
Card snapshot main · updated 2026-10-11
TheDrainFlorist/Qwen3.8-27B-VQ-4.8bpw
- Reference
- bf16 teacher, cached top-64
- Eval data
- three house corpora: prose, public code, literary
- Context
- 12288 tokens (older table: prose-only at 2048)
- Tool
- streamed KL instrument; mlx-lm for older table
Card snapshot main · updated 2026-10-11
thousandml/Qwen3.8-27B-CQ-4bit
- Reference
- original Qwen3.8-27B
- Eval data
- WikiText Raw
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-10-16
timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit
- Reference
- control-bf16 (reasoning-merged BF16 model before abliteration)
- Eval data
- frozen local project benchmarks (harmful-prompt suite, capability macro, code, HumanEval, long-form, MMMU30)
- Context
- not stated
- Tool
- same local harness; MLX 8-bit conversion with mlx-vlm
Card snapshot main · updated 2026-08-25
ToBeStyled/Qwen3.8-27B-ColdFusion-GAIN-Blackwell-DFlash2-Ultra-V1.0
- Reference
- none (absolute perplexity, no FP16 reference); task accuracy checked directly
- Eval data
- perplexity: mixed code/markdown corpus, 24 chunks; GSM8K, HumanEval-50
- Context
- 2048 per chunk, KV f16
- Tool
- llama-perplexity
Card snapshot main · updated 2026-09-11
tooltd/Qwen3.8-27B-GSQ-RCO-BuffedMod-GGUF
- Reference
- original BF16 base model (q38f16baseline.kld)
- Eval data
- wikitext-2-raw wiki.test.raw
- Context
- not stated
- Tool
- llama-perplexity --kl-divergence
Card snapshot main · updated 2026-09-19
tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF
- Reference
- BF16 baseline (q38f16baseline.kld)
- Eval data
- wikitext-2-raw test; EvalPlus HumanEval (164) / MBPP (378) for coding
- Context
- not stated
- Tool
- llama-perplexity --kl-divergence; EvalPlus pass@1 no-thinking, KV cache q4_0
Card snapshot main · updated 2026-09-20
trevest/Qwen3.8-27B-Tetris-GGUF
- Reference
- Q8_0 (measured +0.13% vs BF16 on matched chunks, treated as ~BF16)
- Eval data
- 120 x 512-token chunks of wikitext-2-test + code files held out of calibration set
- Context
- 512-token chunks
- Tool
- llama.cpp perplexity; imatrix 40% bartowski datav3, 48% real code, 12% ChatML tool-calls
Card snapshot main · updated 2026-08-20
Triody/Qwen3.8-27B-NVFP4-AllLinear
- Reference
- Qwen/Qwen3.8-27B bf16, no speculation; also nvidia/Qwen3.8-27B-NVFP4
- Eval data
- GSM8K first 400 test questions; HumanEval 164 problems
- Context
- at most 4096 output tokens
- Tool
- SGLang 0.5.20, greedy, thinking on, strict vs lenient scoring
Card snapshot main · updated 2026-09-30
Twu31/Qwen3.8-27B-AWQ-INT4-MTP-LowLatency
- Reference
- BF16 base Qwen/Qwen3.8-27B (per-token NLL of the BF16 base's own greedy answers under the quantized model)
- Eval data
- 18 conversational prompts (~1.5k-token system prompt + history), 384x2048 calibration for AutoRound
- Context
- not stated
- Tool
- vLLM 0.27.1 on RTX 5090 (fidelity tables); AutoRound 0.14.2, llm-compressor 0.13.0
Card snapshot main · updated 2026-10-07
UgurI/Qwen3.8-27B-Turkish-CPT-GGUF
- Reference
- Base Model Qwen3.8-27B (quant/file of base column not stated)
- Eval data
- Turkish validation and test sets, TurkishMMLU (90-Q stratified), Belebele TR/EN
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-08-23
varshavvvvv/Ternary-CyberX-2-27B-gguf
- Reference
- Qwen3.8-27B FP16
- Eval data
- 14 benchmarks across six skill categories, thinking mode
- Context
- not stated
- Tool
- EvalScope + vLLM on NVIDIA H100
- Card's numbers match TechnoBaptist/Ternary-Bonsai-2-27B-gguf's card (78 of 78 claim lines identical, detected automatically); treated as copied, not as an independent measurement
Card snapshot main · updated 2026-09-22
vrfai/Qwen3.8-27B-FP8
- Reference
- bf16 weights of the same model (measured by author); also Qwen published card figures
- Eval data
- 10 tasks, 13,046 documents, deterministic scorers only (drift table); MathVision testmini and ERQA absolute scores
- Context
- not stated
- Tool
- not stated (vLLM >= 0.27.1 for serving)
Card snapshot main · updated 2026-08-20
vrfai/Qwen3.8-27B-FP8-dynamic
- Reference
- bf16 weights of same model; drift reported as delta vs bf16 with a bf16-vs-bf16 batch-size floor
- Eval data
- 10 tasks, 13,046 documents, deterministic scorers only (AI2D, ChartQA, ScienceQA-IMG, MMStar, OCRBench, CountBenchQA, RealWorldQA, OmniDocBench, MathVision testmini, ERQA)
- Context
- not stated
- Tool
- vLLM >= 0.27.1
Card snapshot main · updated 2026-08-20
vrfai/Qwen3.8-27B-NVFP4
- Reference
- Qwen3.8-27B bf16 weights, same harness, run twice to get a noise floor
- Eval data
- 10 deterministic-scorer tasks, 13,046 documents (AI2D, ChartQA, ScienceQA-IMG, MMStar, OCRBench, CountBenchQA, RealWorldQA, OmniDocBench, MathVision testmini, ERQA)
- Context
- not stated
- Tool
- vLLM on H100 (Marlin path, effectively W4A16); thinking on, greedy
Card snapshot main · updated 2026-08-20
WarlaxZ/Qwen3.8-27B-NInfer-3090
- Reference
- Perplexity: upstream neroued/Qwen3.8-27B-NInfer; KLD: Unsloth Q8_0 GGUF
- Eval data
- perplexity-1m and perplexity-heldout-2026-09 (wikipedia en/zh, arXiv...); KLD on held-out corpus 8 x 4096 tokens per stream
- Context
- 4096-token streams for KLD; int8 KV for perplexity
- Tool
- ninfer-perplexity --quick --kv-dtype int8; llama.cpp b11316 for KLD
Card snapshot main · updated 2026-10-02
WaveCut/Qwen3.8-27B-MLX-4bit-DWQ
- Reference
- pinned 8-bit MLX teacher (not BF16)
- Eval data
- fixed disjoint 64-example holdout
- Context
- rendered length capped at 513 tokens
- Tool
- KL divergence to sparse top-1,024 teacher logits (DWQ training)
Card snapshot main · updated 2026-08-15
wiklif/Qwen3.8-27B-KTopt-GGUF
- Reference
- unsloth/Qwen3.8-27B-GGUF Q8_0 (benchmarks); BF16 (KLD)
- Eval data
- GPQA Diamond, IFBench, LiveCodeBench v6 (629 questions); KLD on code corpus (14 GitHub repos) and prose corpus (arXiv abstracts + Polish Wikipedia)
- Context
- KLD context 4096; benchmark output limit 200k
- Tool
- ik_llama.cpp fork; KLD vs BF16
Card snapshot main · updated 2026-09-15
XpressAI/Qwen3.8-27B-RYS-GGUF
- Reference
- unmodified base Qwen3.8-27B (not BF16 logits)
- Eval data
- 17-prompt reasoning probe, EQ-140, fixed 1,006-case BFCL v4 single-turn manifest across 13 categories
- Context
- not stated
- Tool
- BFCL harness; layer duplication of layers 16-19, no retraining
Card snapshot main · updated 2026-08-15
yeasah/Qwen3.8-27B-exl3
- Reference
- unquantized Qwen3.8-27B
- Eval data
- agentic and coding oriented slices (WildChat, Open-SWE-Traces, documents in user turn, documents as tool results, multilingual, tool-use loops)
- Context
- not stated
- Tool
- excess KLD (minus unquantized model's own rounding noise) and delta perplexity; exllamav3
Card snapshot main · updated 2026-10-12
YourHighnessLA/L0xRE-27b-Low-MTP
- Reference
- native-quant baseline
- Eval data
- internal non-canonical suite, 150 tasks pass@1, thinking on
- Context
- not stated
- Tool
- not stated
Card snapshot main · updated 2026-09-25
Youssofal/Qwen3.8-27B-MTPLX-Bare-Speed
- Reference
- original bf16 Qwen3.8-27B checkpoint
- Eval data
- teacher-forced over 2,389 positions of code, prose, JSON and a multilingual notice (pack quality); own coding battery (Bare Speed KL)
- Context
- not stated
- Tool
- MTPLX 2.11.3
Card snapshot main · updated 2026-09-17
Youssofal/Qwen3.8-27B-MTPLX-Bare-Speed-FP16
- Reference
- bf16 Qwen3.8-27B checkpoint
- Eval data
- 2,389 positions of code, prose, JSON and a multilingual notice
- Context
- not stated
- Tool
- teacher-forced top-1 agreement and KL (MTPLX 2.11.3)
Card snapshot main · updated 2026-09-17
Youssofal/Qwen3.8-27B-MTPLX-Optimized-Quality
- Reference
- bf16 Qwen3.8-27B checkpoint
- Eval data
- 2,389 positions of code, prose, JSON and a multilingual notice; separately a 'coding battery'
- Context
- not stated
- Tool
- teacher-forced top-1 agreement and KL; MTPLX 2.11.3
Card snapshot main · updated 2026-09-22
Youssofal/Qwen3.8-27B-MTPLX-Optimized-Quality-FP16
- Reference
- bf16 Qwen3.8-27B checkpoint
- Eval data
- 2,389 teacher-forced positions of code, prose, JSON and a multilingual notice
- Context
- not stated
- Tool
- teacher-forced top-1 agreement and KL (MTPLX 2.11.3)
Card snapshot main · updated 2026-09-22
Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed
- Reference
- bf16 Qwen3.8-27B checkpoint (teacher-forced)
- Eval data
- 2,389 positions of code, prose, JSON and a multilingual notice
- Context
- not stated
- Tool
- MTPLX 2.11.3
Card snapshot main · updated 2026-09-24
Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16
- Reference
- bf16 Qwen3.8-27B checkpoint
- Eval data
- 2,389 positions of code, prose, JSON and a multilingual notice, teacher-forced
- Context
- not stated
- Tool
- MTPLX 2.11.3
Card snapshot main · updated 2026-09-22
YTan2000/Qwen3.8-27B-TQ
- Reference
- BF16
- Eval data
- GSM8K (1319), MMLU-Pro (291-question subset), GPQA Diamond (198); KLD corpus not stated
- Context
- not stated for KLD; 32K context for speed runs
- Tool
- tqLLM release engine, DFlash2-TQ3 drafter, tq3 KV; KLD on GB10
Card snapshot main · updated 2026-10-14
YTan2000/Qwen3.8-27B-TQ3_4S
- Reference
- none (absolute task scores; no bf16 comparison)
- Eval data
- Hard86 (20 tasks/86 checks), HumanEval/HumanEval+/MBPP/MBPP+ (official scorer pass@1), custom coding/reason-math/tool-call/data-extract/instruction-follow suites
- Context
- not stated
- Tool
- llama.cpp, draft-MTP --spec-draft-n-max 2, reasoning off; evalplus under budget-256
Card snapshot main · updated 2026-10-08
zarigata/FeveQwen-1.0
- Reference
- not stated
- Eval data
- 142 paired prompts tune set (200-pair sealed test not opened)
- Context
- not stated
- Tool
- OBLITERATUS built-in behavioral checks
Card snapshot main · updated 2026-09-24
zedongpeng/Qwen3.8-27B-NInfer-4090
- Reference
- FP32-compute reference over the BF16 checkpoint
- Eval data
- 46,059 next-token predictions in 44 windows, half Wikipedia prose and half CPython source, disjoint from calibration
- Context
- 2,048 tokens per window
- Tool
- KL64 = KL(reference || model) on reference top-64 tokens; NInfer patched engine
Card snapshot main · updated 2026-10-06
zerodigest/Qwen3.8-27B-YMQ-MTP-GGUF
- Reference
- not stated
- Eval data
- WikiText-2 test corpus
- Context
- 4096
- Tool
- llama-perplexity
Card snapshot main · updated 2026-09-02
Zynerji/Ektome-Qwen3.8-27B-PristinelyUncensored
- Reference
- pristine Qwen3.8-27B (before abliteration)
- Eval data
- MMLU-validation loglikelihood accuracy 0-shot; AdvBench refusal (n=100) with keyword classifier
- Context
- not stated
- Tool
- ektome_abliterate
Card snapshot main · updated 2026-08-20
Zynerji/Ektome-Qwen3.8-27B-PristinelyUncensored-GPTQ-MTP
- Reference
- bf16 parent (Ektome-Qwen3.8-27B-PristinelyUncensored)
- Eval data
- cais/mmlu validation, 400 questions, seed 0; refusal compliance n=100 greedy
- Context
- not stated
- Tool
- vLLM 0.27.1 on RTX PRO 6000 (sm_120), single-letter generative answers
Card snapshot main · updated 2026-08-17
Repos without extracted claims (844)
Two different situations, marked in the Status column. Read in full: no quality claims means a reviewer read the whole card and found no quality number about the quant itself. Not yet reviewed means the card has not been read in full yet and may contain claims; it is listed so the catalogue is complete, not as a finding.
| Repository | Status | Format | Weights | Note | Card |
|---|---|---|---|---|---|
| 0bserverx/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| 0xBakeer/TandemLLM-Qwen3.8-27B-EXL3-3.0 | read in full: no quality claims | EXL3 | main | ||
| 0xBakeer/TandemLLM-Qwen3.8-27B-EXL3-3.5 | read in full: no quality claims | EXL3 | main | ||
| 0xBakeer/TandemLLM-Qwen3.8-27B-EXL3-4.0 | read in full: no quality claims | EXL3 | main | ||
| 0xKitkat/Qwen3.8-27B-Uncensored-Aggressive | not yet reviewed | other | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| 0xSojalSec/Qwen3.8-27B-full-Uncensored | not yet reviewed | other | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| 0xSojalSec/Qwen3.8-27B-MLX-2bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| 2531oupp/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| 5oo6w32cu/Qwen3.8-27B-Q8_0-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| a4981678/Qwen3.8-27B-Uncensored-MLX | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| a4lg/Qwen3.8-27B-MTP-ONLY-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| abhishek1kr/Qwen3.8-27B-Uncensored-MLX_demo | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| abhishekchohan/Qwen3.8-27B-AWQ-INT4 | read in full: no quality claims | AWQ | main | ||
| Abiray/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| abood365/Qwen3.8-27B-Uncensored-MLX | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Abyyyyyyyyyyyyyyyy/Qwen3.8-27B-Uncensored-MLX | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| aclava/Qwen3.8-27B-Q3_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| aclava/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| adaliboy/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| adrianmurray/Qwen3.8-27B-Hybrid-Q8Q4 | read in full: no quality claims | other | No quality claims for the quant (card read in full): size and a functional smoke test with speed/acceptance only [L37,L76-80] | main | |
| adrianmurray/Qwen3.8-27B-MLX-6bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| agentionai/Qwen3.8-27B-AP-GGUF | read in full: no quality claims | GGUF | main | ||
| agiws/Qwen3.8-27B-UD-IQ2-S | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| agnosticeng/Qwen3.8-27B-2bit | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| AGW06/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| ahmed-amedo/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| ahmed-amedo/Qwen3.8-27B-OBLITERATED | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| ahmed-amedo/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| AI-conn/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| AiAF/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Card is empty or missing | main | |
| aiartorious/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| AIconjured/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-Q5_NVFP4 | read in full: no quality claims | NVFP4 | No quality claims for the quant (card read in full): size/BPW table only [L49-53]; only perplexity is an imatrix calibration figure, not a quant comparison [L92]; quality section is qualitative [L145-153] | main | |
| aidj2025/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| ailahyhx/Qwen3.8-27B-GGUF-backup | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| airagrp/Qwen3.8-27B-MLX-mxfp8-mixed | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| airagrp/Qwen3.8-27B-MLX-nvfp4-mixed | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Akicou/Qwen3.8-27B-DFlash2-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): draft-model size/BPW tables and speculative acceptance-length (speed) figures quoted from z-lab; no quant fidelity numbers [L26-31,L94-99,L103-113] | main | |
| akopytko/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| aleada/Qwen3.8-27B-W4A16 | read in full: no quality claims | INT4 | No quality claims for the quant (card read in full): only speculative-decoding draft acceptance and speed-up tables plus recipe and size (18.11 GiB) [L77-85, L112-118] | main | |
| aleada/Qwen3.8-27B-W4A16-atlas-tuned | not yet reviewed | INT4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| alex-and-the/Qwen3.8-27B-MLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| alexanderou/Qwen3.8-27B-W4A16-AutoRound-fast-ru | not yet reviewed | AutoRound | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| AlexWortega/Qwen3.8-27B-W4A16-AutoRound | not yet reviewed | AutoRound | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| allhailalona111/Qwen3.8-27B-CRACK-Q4_K_M-MTP-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| alphaduriendur/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| amd/Qwen3.8-27B-fp32-ve-fp32-int4-k_quant-gs128-text-cpu-onnx | not yet reviewed | ONNX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| amd/Qwen3.8-27B-Quark-AWQ-MXFP4 | read in full: no quality claims | MXFP4 | main | ||
| amd/Qwen3.8-27B-w8a8-llmcompressor | read in full: no quality claims | INT8 | main | ||
| amd/Qwen3.8-27B_gpuep_1.0_gpu_128 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| andyjack/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| apertixai/cerno-27b | read in full: no quality claims | other | main | ||
| apriliacloset/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| aquaduck/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| aquaduck/Qwen3.8-27B-MLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Ar4ikov/Qwen3.8-27B-AWQ-W4A16-ASYM | not yet reviewed | AWQ | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Archsloth/Qwen3.8-27B-GGUF | read in full: no quality claims | GGUF | main | ||
| ariel-pillar/Qwen3.8-27B-GGUF_POC | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Arkane-cloud/qwen38-27b-w4a16-autoround | not yet reviewed | AutoRound | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Arki05/Qwen3.8-27B-GGUF-shards | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Arslan2703/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| ArtomYuan/Qwen3.8-27B-ROCmFPX | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| asdhbj/Qwen3.8-27B-FP8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| ashesofthem/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-NVFP4 | read in full: no quality claims | NVFP4 | main | ||
| Astark-40/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Atomic-Germ/Qwen3.8-27B-NPU2 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Atomic-Germ/Qwen3.8-Whittle-MoE-27B-A17.8B-GGUF | read in full: no quality claims | GGUF | main | ||
| ATOMIKMN/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| atypical-chai/Qwenmodel1st | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| AutomatosX/AX-Qwen3.8-27B-CUDA-AXQ-NVFP4-W4A16-MTP | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| AutomatosX/AX-Qwen3.8-27B-MLX-AXQ-MXFP4-MTP | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| AutomatosX/AX-Qwen3.8-27B-MLX-AXQ-MXFP8-MTP | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| autotrust/Qwen3.8-27B-NVFP4-DGX-SPARK | read in full: no quality claims | NVFP4 | No quality claims for the quant (card read in full): speed tables only [L63-80]; weights bit-exact identical to upstream nvidia repo [L28], no quality numbers | main | |
| AvatarAndrey/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Avifenesh/Qwen3.8-27B-NVFP4-MTP-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): draft acceptance rate and decode speed only, no quant quality metric [L31-36, L79-87, L132-136, L294-301] | main | |
| Axel0689/Qwen3.8-27B-Q3_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| axiomofmind/Qwen3.8-27B-NVFP4-W4A16-MLP-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): file sizes, layout table and one DFlash2 speed/acceptance run only [L27-32, L51-57, L133] | main | |
| axum360/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-Q4KS-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Aydge/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| ayqzs/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| AzatAI/Qwen3.8-27B-FP8-dynamic | read in full: no quality claims | FP8 | No quality claims for the quant (card read in full): size and latency/memory tables only [L76-88]; card says no evaluation suite has been run [L133] | main | |
| azukivc/Qwen3.8-27B-Abliterated-SFT | not yet reviewed | other | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| azukivc/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| barrydeen/Qwen3.8-27B-AWQ-4bit | read in full: no quality claims | AWQ | No quality claims for the quant (card read in full): MTP acceptance/throughput only [L96-105]; card states no formal benchmark suite was run [L130] | main | |
| batiai/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| ben0112/Qwen-Qwen3.8-27B-MTPLX | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Bendyline/Qwen3.8-27B-mtp-drafter-mlx-4bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| bgdfyymbw/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| bielquants/Ternary-Bonsai-2-27B-gguf | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| biMEMO/Qwen3.8-27B-int4-AutoRound | read in full: no quality claims | AutoRound | No quality claims for the quant (card read in full): throughput tables and MTP draft acceptance only [L52-59, L74, L106]; quality asserted qualitatively (needle test, no quality drop) with no numeric score [L77-83] | main | |
| blackfan23/Dirk-Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| bluehawana/Qwen3.8-27B-Q8-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| BnhHo/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| BoldingBuilds/Qwen3.8-27B-Abliterated-ThinkFix-GGUF | read in full: no quality claims | GGUF | main | ||
| BoldingBuilds/Qwen3.8-27B-Reasoning-Termination-Fix-GGUF | read in full: no quality claims | GGUF | main | ||
| born2bewild/Qwen3.8-27B-W4A16-AutoRound-fast | not yet reviewed | AutoRound | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| bowmanslayer/kev-27b-W4A16 | not yet reviewed | INT4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| bowmanslayer/Qwen3.8-27B-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): size/pp512/tg128/context tables [L45-52] and MTP draft tok/s and acceptance [L131-163]; quality only described qualitatively [L59-61] | main | |
| bowmanslayer/Qwen3.8-27B-W4A16-vision-mtp | not yet reviewed | INT4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| br1nz/Qwen3.8-27B-Just-MTP-NVFP4.gguf | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| bramburn/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| breach-irate-punch/Qwen3.8-27B-mlx-4Bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Brian6145/Qwen3.8-27B-Imatrix-Mtp-1m | read in full: no quality claims | other | No quality claims for the quant (card read in full): MTP draft acceptance and speed tables only [L82-96] | main | |
| brianbellx/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF-copy | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Brooooooklyn/Qwen3.8-27B-MXFP4-mlx | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Brooooooklyn/Qwen3.8-27B-NVFP4-mlx | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| bsaleh03/Qwen3.8-27B-ASCII-Condensed | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| bwn2000/moeme-27b | read in full: no quality claims | other | main | ||
| bytehackr/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| ByteOtter/Qwen3.8-27B-TAK-Reasoning-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| byteshape/Qwen3.8-27B-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): quality vs TPS shown only as an image chart [L49-56]; text tables give size, bits/weight and TPS only [L62-68, L104-110] | main | |
| bytkim/Qwen3.8-27B-pi-FP8 | read in full: no quality claims | FP8 | No quality claims for the quant (card read in full): benchmark results (Terminal-Bench, GPQA, SciCode) appear only as images and compare a Pi fine-tune vs Base, no numeric quant-quality values in text [L37, L59-83] | main | |
| bytkim/Qwen3.8-27B-pi-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): text tables are MTP head sizes, decode speedups and KV-cache sizes only [L126-130, L219-227]; quality benchmarks are in images only [L40, L76] | main | |
| cafepm/Qwen3.8-27B-Q3Q4-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| callgg/qwen3.8-27b-gguf | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Callick99/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| canhdu/Qwen3.8-27B-IQ3_S-FFN-IQ4_XS-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| catplusplus/Qwen3.8-27B-heretic-ara-NVFP4-MTP | read in full: no quality claims | NVFP4 | No quality claims for the quant (card read in full): only drafter acceptance rate (speed) and anecdotal use note [L8-10] | main | |
| Charlesies/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Chopinab/Qwen3.8-27B-Q2_K-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| christian-bick/EduGraph-Classifier-Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| chriszib/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Chungulus/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Chungulus/Qwen3.8-27B-IMATRIX-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Chungulus/Qwen3.8-27B-IQ4_NL-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Chungulus/Qwen3.8-27B-IQ4_XS-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Chungulus/Qwen3.8-27B-MLX-5bit | read in full: no quality claims | MLX | main | ||
| Chungulus/Qwen3.8-27B-MLX-6bit | read in full: no quality claims | MLX | main | ||
| Chungulus/Qwen3.8-27B-MLX-8bit | read in full: no quality claims | MLX | main | ||
| Chungulus/Qwen3.8-27B-MLX-MXFP4 | read in full: no quality claims | MLX | main | ||
| Chungulus/Qwen3.8-27B-MTP-GGUF | read in full: no quality claims | GGUF | main | ||
| Chungulus/Qwen3.8-27B-Q2_K-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Chungulus/Qwen3.8-27B-Q3_K_L-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Chungulus/Qwen3.8-27B-Q3_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Chungulus/Qwen3.8-27B-Q3_K_S-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Chungulus/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Chungulus/Qwen3.8-27B-Q4_K_S-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Chungulus/Qwen3.8-27B-Q5_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Chungulus/Qwen3.8-27B-Q5_K_S-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Chungulus/Qwen3.8-27B-Q6_K-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Chungulus/Qwen3.8-27B-Q8_0-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Ckary/Qwen3.8-27B-Uncensored-MLX | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| cloudnathan5/Qwen3.8-27B-NVFP4a16-AWQ | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Codydaoust/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| cole17e/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| compute1/Qwen3.8-27B-W8A8-INT8 | not yet reviewed | INT8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| CorVous/Qwen3.8-27B-heretic-ara-GGUF | read in full: no quality claims | GGUF | main | ||
| CrossNow/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| CrossNow/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| cryptonaut/Qwen3.8-27B-Uncensored-Aggressive | not yet reviewed | other | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| ctz168/Qwen3.8-27B-GGUF-q4km | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| cvgro/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| cxc-ryan/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| cyankiwi/Qwen3.8-27B-AWQ-BF16-INT4 | read in full: no quality claims | AWQ | No quality claims for the quant (card read in full): benchmark tables [L98-221, L229-259] are the upstream Qwen3.8-27B base model results, not measurements of this quant; only model size stated [L33] | main | |
| cyankiwi/Qwen3.8-27B-AWQ-FP8 | read in full: no quality claims | AWQ | No quality claims for the quant (card read in full): only model size 31.28 GB [L34]; benchmark tables [L112-260] are the upstream Qwen3.8-27B model card, not quant measurements | main | |
| cyankiwi/Qwen3.8-27B-AWQ-INT4 | read in full: no quality claims | AWQ | No quality claims for the quant (card read in full): benchmark tables [L99-222, L230-260] are the upstream Qwen3.8-27B base model results, not measurements of this quant; only model size stated [L34] | main | |
| CyberGamer/Qwen3.8-27B-UD-IQ2_S-RouteCache-GGUF | read in full: no quality claims | GGUF | main | ||
| Cyrile/Qwen3.8-27B-FP8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| dabeet24/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| dancinlab/Qwen3.8-27B-Uncensored-FP8 | not yet reviewed | FP8 | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| dancinlab/Qwen3.8-27B-Uncensored-MLX | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| darioooooo0o/Qwen3.8-27B-GPTQ-Int4-asym | read in full: no quality claims | GPTQ | main | ||
| darkbit1001/Qwen-3.8-27B-exl3-4.500bpw-hb6 | read in full: no quality claims | EXL3 | No quality claims for the quant (card read in full): only the original Qwen3.8-27B model-card benchmark tables for the base model, not the quant [L87-239] | main | |
| darkbit1001/Qwen3.8-27B-exl3-4.00bpw-hb8 | read in full: no quality claims | EXL3 | No quality claims for the quant (card read in full): only quant settings [L18-26] and the upstream Qwen3.8-27B original model-card benchmark tables [L87-212, L220-251], which describe the base model not this quant | main | |
| datamart99/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| DavidAU/Qwen3.8-27B-Turbo-Brilliance-50X-Reasoning-Instruct-NEO-MAX-MTP-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): no quant quality metrics; only usage guidance and sample generations [L194-214, L717-724] | main | |
| davisliu/Qwen3.8-27B-MXFP4-VL-MTP | not yet reviewed | MXFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| dbrasdasilva/Qwen3.8-27B-FP8-MTP | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| dbrasdasilva/Qwen3.8-27B-NVFP4-MTP | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| dbrasdasilva/Qwen3.8-27B-Text-NVFP4-MTP | read in full: no quality claims | NVFP4 | No quality claims for the quant (card read in full): KV-cache ceilings, context and decode tok/s only; no accuracy or divergence numbers [L58-68, L107-141] | main | |
| ddalcu/Qwen3.8-27B-MLX-Serve-4bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| ddalcu/Qwen3.8-27B-MLX-Serve-8bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| dealignai/Qwen3.8-27B-UNCENSORED-GGUF | read in full: no quality claims | GGUF | main | ||
| dendriteholdings/Ingot-genesis-27B-FP8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| deresolution/Qwen3.8-27B-mxfp4 | not yet reviewed | MXFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| derikn/Qwen3.8-27B-AWQ-INT4 | not yet reviewed | AWQ | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| desva0/Qwen3.8-27B-Uncensored-W4A16 | not yet reviewed | INT4 | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Deuiz/Qwen3.8-27B-Q3_K_S-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| devan-carlin/Qwen3.8-27B-int4-AutoRound | not yet reviewed | AutoRound | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| developerjeremylive/Ternary-Bonsai-2-27B-gguf-etheroi | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| DevQuasar/Qwen.Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| dfischermittwald/Qwen3.8-27B-NVFP4-DFlash2 | read in full: no quality claims | NVFP4 | No quality claims for the quant (card read in full): decode-throughput table only [L93-102]; fidelity is a pass count (OCR 36/36) [L167-168] and truncated KL whose numbers live in KLD.md, not the card [L170-174] | main | |
| DGM0N/Qwen3.8-27B-exl3-3.125bpw | read in full: no quality claims | EXL3 | No quality claims for the quant (card read in full): card is the upstream Qwen3.8-27B base-model card; benchmark tables [L60-224] are base-model scores, no number about this quant | main | |
| dhruv1672/spartan-studybuddy-qwen3.8-27b-nvfp4 | read in full: no quality claims | NVFP4 | No quality claims for the quant (card read in full): speculative decoding throughput/TTFT/acceptance table only; fine-tune eval_loss is not a quant-quality comparison [L76-95,L61] | main | |
| DhruvalLabs/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Didkk/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| DimInfer/Qwen3.8-27B-Dspark-v1 | read in full: no quality claims | other | No quality claims for the quant (card read in full): speculative-decoding drafter; decode speed, accepted length and speedup tables only [L36-46, L58-63, L108-115] | main | |
| Distillio/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Donee/Qwen3.8-27B-Uncensored-MLX | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Dongwei18416/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| DonnyFlo85/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| douyamv/Qwen3.8-27B-FP8 | read in full: no quality claims | FP8 | No quality claims for the quant (card read in full): only qualitative 'near-original accuracy' wording and size (~27 GB) / VRAM reduction (~50%) [L36-37, L42] | main | |
| douyamv/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| dowands/NeoJev-Qwen3.8-27B-L56-MLX | read in full: no quality claims | MLX | main | ||
| dpstill/Qwen3.8-27B-GDNMix-NVFP4-NInfer | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| DreamFoundries/Qwen3.8-27B-6bit | read in full: no quality claims | other | No quality claims for the quant (card read in full): size and bits-per-weight statement only [L14] | main | |
| DreamFoundries/Qwen3.8-27B-8bit | read in full: no quality claims | other | No quality claims for the quant (card read in full): size and bits-per-weight only [L14] | main | |
| dudeman2512/Qwen3.8-27B-FP8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| dudeman2512/Qwen3.8-27B-int4 | not yet reviewed | INT4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| ediuedddward/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| edp1096/Huihui-Qwen3.8-27B-abliterated-FP8 | not yet reviewed | FP8 | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| eebyp/Qwen3.8-27B-GGUF-Q4_K_M | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| eemin/Qwen3.8-27B-NVFP4-MTP-2x16GB | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| eemin/Qwen3.8-27B-ROCmI4-MTP-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| el4/Qwen3.8-27B-ONYX-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Elmerfudge/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| EnclaveHost/qwen3.8-27b-mtp-gguf | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): GGUF header geometry and serving-size arithmetic only [L43-57, L104-117]; no fidelity or benchmark numbers | main | |
| EnclaveHost/qwen3.8-27b-mtp-q8-vl-gguf | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): decode speed table and short smoke checks only, explicitly not an accuracy evaluation [L116-119, L124-128] | main | |
| eniairaph07/Qwen3.8-27b-FABLE-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): benchmark tables [L29-62] are model-level scores with no per-quant measurement; GGUF table is sizes only [L68-73] | main | |
| erobey/Qwen3.8-27B-mlx-4bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| erobey/Qwen3.8-27B-mlx-8bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| exeterminal/Exe-Core-Dynamic-V1-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): file type/bits/size table only [L74-85]; no per-quant quality numbers shown (evaluation section L111-123 concerns the fine-tune, no figures in candidates) | main | |
| FAISALFAZALHUSSAIN/Qwen3.8-27B-FP8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Fancp/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Farggin/Lobo-Qwen3.8-27B-GSQ-RCO-IQ3_S-MTP-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): artifact size and provenance only; no quality numbers [L18-39] | main | |
| Fastiraz/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| fcmeyer/Qwen3.8-27B-MLX-oQ4e-mtp | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): speed/size/draft-acceptance only; no fidelity or accuracy numbers [L19, L26] | main | |
| FenomAI/Qwen3.8-27B-AWQ-INT4 | read in full: no quality claims | AWQ | No quality claims for the quant (card read in full): only model size (21.02 GB) in header table [L34]; benchmark tables [L96-260] are the official base-model Qwen3.8-27B results, not this quant | main | |
| FenomAI/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Fhjdjdbxhshsh/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| FlatFootInternational/Qwen3.8-27B-mlx-6bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| FlatFootInternational/Qwen3.8-27b-MTPLX-Q6G64 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Flow-wolF69/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| FokusDK/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Fr4nkSt3in/Qwen3.8-27B-BF16-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| FreedomAISVR/Qwen3.8-27B-MXFP4-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| FreedomAISVR/Qwen3.8-27B-NVFP4-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| FreeMangione221/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Frozenlock/Qwen3.8-27B-int4-AutoRound | read in full: no quality claims | AutoRound | No quality claims for the quant (card read in full): only size, MTP draft acceptance and speed stated; no accuracy/KL/perplexity numbers [L18-20, L34-42] | main | |
| fyqfy/Qwen3.8-27B-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): copied Unsloth upstream card with no tables or measured values; only an unquantified marketing claim of better top-1 accuracy [L35] and base-model description | main | |
| gabmiranda/Qwen3.8-27B-oQ2.5e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| gabmiranda/Qwen3.8-27B-oQ2.7e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| gabmiranda/Qwen3.8-27B-oQ2e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| gabmiranda/Qwen3.8-27B-oQ3.5e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| gabmiranda/Qwen3.8-27B-oQ3e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| gabmiranda/Qwen3.8-27B-oQ4e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| gabmiranda/Qwen3.8-27B-oQ5e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| gabmiranda/Qwen3.8-27B-oQ6e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| gabmiranda/Qwen3.8-27B-oQ8e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| GalaxyS2Gordon/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| gaunernst/Qwen3.8-27B-MXFP6 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| GCSA-AiLab/Qwen3.8-27B-Abliterated-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| geezyforeezy/Qwen3.8-27B-OBLITERATED | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| ggh5454/Qwen3.8-27B-MLX-4bit | read in full: no quality claims | MLX | main | ||
| ggh5454/Qwen3.8-27B-MLX-8bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| gghfez/Qwen3.8-27B-Q5_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| ggml-org/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| ghgdhhsacd/Qwen3.8-27B-Ridge-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| gihakkk/Qwen3.8-27B-GGUF-test | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| giocom/Qwen3.8-27B-MXFP4-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Gokugamer/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| GoldfingerCH/Qwen3.8-27B-abliterated-NVFP4 | not yet reviewed | NVFP4 | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| GotoAI-Inc/Qwen3.8-27B-W4A16 | not yet reviewed | INT4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| gpillon/Qwen3.8-27B-nvfp4full-dflash2-NInfer | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| GreenBitAI/Qwen3.8-27B-4bit | read in full: no quality claims | other | No quality claims for the quant (card read in full): decode speed and draft acceptance tables only [L28-48]; claims output unchanged by draft head but gives no quality number [L19-20] | main | |
| grimoni/Qwen3.8-27B-SSMFIX-UD-Q4_K_XL-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| groxaxo/Qwen3.8-27B-3bit-MLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| groxaxo/Qwen3.8-27B-abliterated-GPTQ-Pro | read in full: no quality claims | GPTQ | No quality claims for the quant (card read in full): MTP speed tables only [L30-35, L120-123, L143-149]; bf16 KV 'no measurable quality penalty' is unquantified [L155] | main | |
| groxaxo/Qwen3.8-27B-Selective-AWQ-A45 | not yet reviewed | AWQ | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| groxaxo/Qwen3.8-27B-UD-IQ4_XS-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| gshe-x2333/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Guile/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| HaberstrohSystems/Qwen3.8-27B-int4-g128-sym-autoround | not yet reviewed | AutoRound | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Hagwell/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Hagwell/Ternary-Bonsai-2-27B-gguf | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| hailun-core/Qwen3.8-27B-mlx-6Bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| hailun-core/Qwen3.8-27B-mlx-8Bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| halt95/Qwen3.8-27B-W4A16-Merlin | not yet reviewed | INT4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| hammadRahman/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| hanchen1145/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| HandleCoding/Qwen3.8-27B-nvfp4-NInfer | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| hardcoremoore/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| hariharanv04/Qwen3.8-27B-mlx-2Bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| harmontr/Qwen3.8-27B-Q2_K-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): size/BPW and MTP speed tables only [L42-60, L92-97, L167] | main | |
| Hefen2959512/Qwen3.8-27B-ZeroRefusal-UD-IQ4_XS-MTP-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Hirt556/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| hitsfmdj/Qwen3.8-27B-GGUF-4.2BPW-16GB | read in full: no quality claims | GGUF | main | ||
| Hjx2/Qwen3.8-27B-Q4_K_S-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| hoborific/Qwen3.8-27B-W8A16-FP8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| holooo/Qwen3.8-27B-Q2_K-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Honkware/Qwen3.8-27B-exl3-3.8bpw | not yet reviewed | EXL3 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Honkware/Qwen3.8-27B-exl3-4.5bpw | not yet reviewed | EXL3 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| htrhdjy/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| huake/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| hugobugo34/Qwen3.8-27B-ROCmFPX-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| hwkranger/Qwen3.8-27B-heretic-ara-NVFP4 | not yet reviewed | NVFP4 | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| ibsophobia/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| iceDonkey/Qwen3.8-27B-FP8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| id-2/rox-r1-gguf | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| ig1/Qwen3.8-27B-NVFP4-FP8KV-MTP | read in full: no quality claims | NVFP4 | No quality claims for the quant (card read in full): calibration description and MTP draft acceptance (~56% with 3 tokens) only, no quality metric vs a reference [L31-36, L92] | main | |
| igor255/Qwen3.8-27B-AWQ-INT4-H8-MTP4 | not yet reviewed | AWQ | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| igor255/Qwen3.8-27B-DFlash2-EXL3-4.00bpw | read in full: no quality claims | EXL3 | No quality claims for the quant (card read in full): speculative-decoding acceptance length and throughput tables only [L81-127]; no quality number for the quant | main | |
| igorvibes/Qwen3.8-27B-UD-Q4_K_XL-AWQ-MTP-mlx | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| igorvibes/Qwen3.8-27B-UD-Q5_K_XL-AWQ-MTP-mlx | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| igorvibes/Qwen3.8-27B-UD-Q6_K_XL-AWQ-MTP-mlx | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| igroc656/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| ikantkode/Qwen3.8-27B-AWQ-vd | read in full: no quality claims | AWQ | No quality claims for the quant (card read in full): no quality number visible in the digest; only size and AWQ settings [L21,L38-39]; the 'Measured performance' table [L23-34] contents are not shown so could not be mapped | main | |
| ilyakam/Geer-Qwen3.8-27B-4-8bit-MLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| ilyakam/Geer-Qwen3.8-27B-6bit-MLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| immiq/Qwen3.8-27B-Orinfer-E8P-Q4A8 | read in full: no quality claims | other | No quality claims for the quant (card read in full): size and compatibility only; card says full LLM benchmarks and BF16 quality evaluation are outside scope [L16, L22, L49] | main | |
| impacte/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| IMUGLYHUH/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| incept5/Qwen3.8-27B-MLX-4bit | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): only a qualitative smoke test and peak memory [L48-100]; benchmark tables [L173-322] are the upstream Qwen3.8-27B model card, not quant measurements | main | |
| incept5/Qwen3.8-27B-MLX-8bit | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): only qualitative smoke test [L46-101] and size/memory notes; benchmark tables [L157-321] are the upstream base-model scores, not this quant | main | |
| incept5/Qwen3.8-27B-MLX-bf16 | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): bf16 cast with no quantization; smoke test not run on bf16 [L51-57]; only the upstream Qwen3.8-27B model benchmarks are pasted [L137-301] | main | |
| incoai/Qwen3.8-27B-DFlash2-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Inferact/Qwen3.8-27B-MXFP4 | not yet reviewed | MXFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Inferact/Qwen3.8-27B-NVFP4 | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| inference-optimization/Qwen3.8-27B-DSpark-FP8-BLOCK | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| inference-optimization/Qwen3.8-27B-DSpark-FP8-DYNAMIC | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| inference-optimization/Qwen3.8-27B-DSpark-Gauss-FP8-W8A8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| inference-optimization/Qwen3.8-27B-DSpark-Gauss-NVFP4-W4A4 | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| inference-optimization/Qwen3.8-27B-DSpark-GPTQ-Gauss-NVFP4-W4A4 | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| inference-optimization/Qwen3.8-27B-DSpark-GPTQ-IMatrix-Gauss-NVFP4-W4A4 | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| inference-optimization/Qwen3.8-27B-DSpark-GPTQ-IMatrix-PerfectBlend-NVFP4-W4A4 | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| inference-optimization/Qwen3.8-27B-DSpark-GPTQ-PerfectBlend-NVFP4-W4A4 | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| inference-optimization/Qwen3.8-27B-DSpark-PerfectBlend-FP8-W8A8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| inference-optimization/Qwen3.8-27B-DSpark-PerfectBlend-NVFP4-W4A4 | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Infermatic/Hemmingway-1-FP8-Dynamic | read in full: no quality claims | FP8 | No quality claims for the quant (card read in full): benchmark charts (CommunicationBench, EQ-Bench 4, StoryBench) are images evaluating the Hemmingway-1 model itself, not the quant; no numeric cells [L39-95] | main | |
| Intel/Qwen3.8-27B-bpw2.8-AutoRound | read in full: no quality claims | AutoRound | main | ||
| Intel/Qwen3.8-27B-q2ks-AutoRound | not yet reviewed | AutoRound | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Intel/Qwen3.8-27B-q4km-AutoRound | not yet reviewed | AutoRound | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Israeli-AI/Qwen3.8-27B-MTP-W4A16-VOLTA-Ampere | read in full: no quality claims | INT4 | No quality claims for the quant (card read in full): size and MTP acceptance-length speed table only [L34-43] | main | |
| ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF | read in full: no quality claims | GGUF | main | ||
| Ivanuk4PDA/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| ixdx/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| JackAgentLead/Jack-3.8-27B-Coder-16GB-VRAM | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| JackAgentLead/Jack-3.8-27B-V2-VL-16GB-VRAM | read in full: no quality claims | other | No quality claims for the quant (card read in full): narrative card, no quant quality numbers; states V2-specific evaluation treated separately and V1 results do not transfer [L599,L733] | main | |
| JackBinary/Qwen3.8-27B-GGUF-ROCMFPX | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| jacobcarajo/Qwen3.8-27B-Q5_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| jaehyeono/Qwen3.8-27B-W3A16-GPTQ | not yet reviewed | GPTQ | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| jakeatx/Qwen3.8-27B-ATX-IQ4_XS-M-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): speed/size/context tables only [L47-55]; quality only 'expected between Q3_K_XL and Q4_K_M', comparison 'in progress' [L55] | main | |
| jalbrethsen-highflame/Qwen3.8-27B-NPU2 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Jaredhollins/Qwen3.8-27B-OBLITERATED | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| jaromer/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| JasonW2025/Qwen3.8-27B-NVFP4-W4A16 | read in full: no quality claims | NVFP4 | main | ||
| jchonde/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| jenerallee78/Qwen3.8-27B-Abliterated-SFT | read in full: no quality claims | other | main | ||
| Jerem62/Qwen3.8-27B-Q8_0-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| jesusoctavioas/Qwen3.8-27B-mlx-2Bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| jesusoctavioas/Qwen3.8-27B-mlx-3Bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| jesusoctavioas/Qwen3.8-27B-mlx-4Bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| jesusoctavioas/Qwen3.8-27B-mlx-8Bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| ji-farthing/Qwen3.8-27B-ik-llama-GGUF | read in full: no quality claims | GGUF | main | ||
| jj33512/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| jksdfosd/Qwen3.8-27B-INT4-both-mtp | not yet reviewed | INT4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| jksdfosd/Qwen3.8-27B-INT4-mtpq | not yet reviewed | INT4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| jmarceno/Qwen3.8-27B-DFlash2-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): draft model GGUF; only file sizes and decode speed/acceptance/prefill tables [L45-49, L135-138]; upstream acceptance lengths are task figures for the draft, not quant quality [L155-157] | main | |
| johninthepool/Qwen3.8-27B-MTPLX-8bit | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| johninthepool/Qwen3.8-27B-VL-mlx-4bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| johnnyeric/Qwen3.8-27B-Uncensored-NVFP4 | read in full: no quality claims | NVFP4 | main | ||
| JonathanColetti/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| josecjr23232/gen-llm-qwen38 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| joydle/Qwen3.8-27B-Arf-GGUF | read in full: no quality claims | GGUF | main | ||
| joyfox/JoyFox-Reverie-27B | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| joyfox/Qwen3.8-27B-Uncensored-JoyFox-Aggressive | not yet reviewed | other | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Jundot/Qwen3.8-27B-oQ2.5e-fp16-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Jundot/Qwen3.8-27B-oQ2.5e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Jundot/Qwen3.8-27B-oQ2e-fp16-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Jundot/Qwen3.8-27B-oQ3.5e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Jundot/Qwen3.8-27B-oQ3e-fp16-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Jundot/Qwen3.8-27B-oQ3e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Jundot/Qwen3.8-27B-oQ4e-fp16-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Jundot/Qwen3.8-27B-oQ4e-mtp | read in full: no quality claims | other | main | ||
| Jundot/Qwen3.8-27B-oQ6e-fp16-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Jundot/Qwen3.8-27B-oQ6e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Jundot/Qwen3.8-27B-oQ8e-fp16-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Jundot/Qwen3.8-27B-oQ8e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| k1lliiviie/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| KaedeTai/dflash2-mlx-quantized-draft | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): speed/size/acceptance tables only [L37-41, L49-53, L61-65] | main | |
| KAISLMR1/Qwen3.8-27B-Uncensored-MLX | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| kaitchup/Qwen3.8-27B-GSQ-INT2-10e | read in full: no quality claims | other | No quality claims for the quant (card read in full): hyperparameter table only [L14-28]; quality only described qualitatively ('significantly underperform') [L12] | main | |
| kaitchup/Qwen3.8-27B-GSQ-INT2-INT3-10e | read in full: no quality claims | other | No quality claims for the quant (card read in full): hyperparameter table only; qualitative 'significantly underperform' with no number [L12-28] | main | |
| kaitchup/Qwen3.8-27B-GSQ-INT2-INT3-20e | read in full: no quality claims | other | No quality claims for the quant (card read in full): hyperparameter table only [L14-28]; text says only that it significantly underperforms the original, with no number [L12] | main | |
| kaitchup/Qwen3.8-27B-GSQ-INT3-10e | read in full: no quality claims | other | No quality claims for the quant (card read in full): hyperparameter table only [L14-28]; prose says models significantly underperform the original without a number [L12] | main | |
| kaitchup/Qwen3.8-27B-hum-quality-MTP-19.2GB | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Kaktus4048/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Kamnsv/Qwen3.8-27B-AWQ-INT4 | not yet reviewed | AWQ | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| kancilbersayap/Qwen3.8-27B-Uncensored-MLX | read in full: no quality claims | MLX | main | ||
| karlsencoin/Karlsen-Chat-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| karnodel/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| KasugaiSakura/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| KasugaiSakura/Qwen3.8-27B-Uncensored-JonathanColetti-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| kdercksen/Kev-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Kecven/Qwen3.8-27B-MTPLX-Q4 | read in full: no quality claims | other | No quality claims for the quant (card read in full): speed and MTP draft-acceptance table only [L46-51] | main | |
| Kecven/Qwen3.8-27B-MTPLX-Q4-Dynamic | read in full: no quality claims | other | No quality claims for the quant (card read in full): speed and MTP acceptance tables only [L55-L60, L74-L79]; no quality metric vs reference | main | |
| Kecven/Qwen3.8-27B-MTPLX-Q8 | read in full: no quality claims | other | No quality claims for the quant (card read in full): speed and MTP draft acceptance only [L46-51, L61-71]; Q4 vs Q8 table is throughput [L75-85] | main | |
| kedric/Qwen3.8-27B-MR-GPTQ-NVFP4-GB10 | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| KENIC-1/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| kenotron/Qwen3.8-27B-mlx-4Bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| keXjos/Qwen3.8-27B-mlx-2Bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| khalifah66362666/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| KikoCis/Qwen3.8-27B-GGUF | read in full: no quality claims | GGUF | main | ||
| kingjones777/Qwen3.8-27B-ROCmFPX-Q8_0-AGENT-GGUF | read in full: no quality claims | GGUF | main | ||
| kingjones777/Qwen3.8-27B-ROCmFPX-Q8_0-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): speed/size table only [L60-64]; card states no perplexity and no quality A/B vs BF16 [L140-141]; only a tools 7/7 functional check [L15] | main | |
| kingjones777/Qwen3.8-27B-Uncensored-ROCmFP4-STRIX-MTP-GGUF | read in full: no quality claims | GGUF | main | ||
| KingKassell/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| kirillbilchenko/Qwen3.8-27B-MXFP4-MxWave-GGUF | read in full: no quality claims | GGUF | main | ||
| KissMyShinyArse/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| kizuna-intelligence/Qwen3.8-27B-int4-qep | not yet reviewed | INT4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Krasnopjorovs/Qwen3.8-27B-Imatrix-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): MTP speed/acceptance table [L39-45] and size table only [L66-79]; no quality number for any quant | main | |
| KridgeDookie/Qwen3.8-27B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS | read in full: no quality claims | other | No quality claims for the quant (card read in full): BF16 checkpoint; only refusal/coherence screens, no quant fidelity numbers [L27,L41] | main | |
| l2dy/Qwen3.8-27B-MXFP4-mlx | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lactroiii/Qwen3.8-27B-FP8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lactroiii/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| laladream/Qwen3.8-27B-Uncensored-MLX | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| LanguaMan/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Launch80/Qwen3.8-27B-PARO-MXFP4 | read in full: no quality claims | MXFP4 | main | ||
| ldov/Ternary-Bonsai-2-27B-gguf | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| LeeMantras/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| lemonyins/Qwen3.8-27B-ULTIMATE-UNCENSORED-MTP-IQ4-GGUF-16GB | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): memory and speed tables only; card says perplexity figures are carried over and pending verification [L57-73, L86] | main | |
| LeonaQuantum/leonaq-1 | read in full: no quality claims | other | main | ||
| lesa80/Qwen3.8-27B-MLX-4bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| letechlead/Huihui-Qwen3.8-27B-Abliterated-INT4-W4A16-AutoRound | not yet reviewed | AutoRound | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| lfitoto/Qwen3.8-27B-Uncensored-NVFP4 | not yet reviewed | NVFP4 | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| LibertAIDAI/Qwen3.8-27B-NVFP4-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): file sizes and speed tables only (llama-bench, batched-bench) [L33-38, L48-63]; no KLD, PPL or benchmark scores; calibration described but vision/multilingual quality explicitly not measured [L85-86] | main | |
| LibertAIDAI/Qwen3.8-27B-NVFP4-MTP-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): file-size table [L35-41] and MTP throughput/acceptance tables [L53-73] only; caveats state vision/multilingual/code not evaluated [L103-104] | main | |
| Litwein/Qwen3.8-27B-oQ4e-DWQ-MTP-MLX | read in full: no quality claims | MLX | main | ||
| livebylord/Qweblethos-v1-GGUF-EXPEREMENTIAL | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): held-out loss/perplexity compares base vs LoRA fine-tune on 128 examples, not any quant's fidelity [L55-62]; Q4_K_M file has only a smoke test [L70] | main | |
| lmstudio-community/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lmstudio-community/Qwen3.8-27B-MLX-4bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lmstudio-community/Qwen3.8-27B-MLX-5bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lmstudio-community/Qwen3.8-27B-MLX-6bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lmstudio-community/Qwen3.8-27B-MLX-8bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| local-inference-lab/Qwen3.8-27B-NVFP4-QAD | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| localweights/Qwen3.8-27B-MTP-IMAT-IQ4_XS-Q8nextn-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lokeshe09/Qwen3.8-27B-FP8-Dynamic | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lokeshe09/Qwen3.8-27B-INT8 | not yet reviewed | INT8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lokeshe09/Qwen3.8-27B-W4A16 | not yet reviewed | INT4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| LoveGlaze/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lstari/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| lstari/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF-q4kp | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Lucebox/Qwen3.8-27B-DFlash2-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lukaskremla/Qwen3.8-27B-2bit-MLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lukaskremla/Qwen3.8-27B-2bit-MLX-TextOnly | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lukaskremla/Qwen3.8-27B-3bit-MLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lukaskremla/Qwen3.8-27B-3bit-MLX-TextOnly | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lukaskremla/Qwen3.8-27B-4bit-MLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lukaskremla/Qwen3.8-27B-4bit-MLX-TextOnly | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lukaskremla/Qwen3.8-27B-5bit-MLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lukaskremla/Qwen3.8-27B-5bit-MLX-TextOnly | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lukaskremla/Qwen3.8-27B-6bit-MLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lukaskremla/Qwen3.8-27B-6bit-MLX-TextOnly | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lukaskremla/Qwen3.8-27B-8bit-MLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lukaskremla/Qwen3.8-27B-8bit-MLX-TextOnly | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| lumos12/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| LunarMoonar/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| LunarMoonar/Huihui-Qwen3.8-27B-abliterated-GGUF-2 | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| LunarMoonar/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| LunarMoonar/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF-6-bit | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): size and effective bits only [L24, L61-62]; card explicitly claims no numerical parity with GGUF [L32] | main | |
| maccelerate/Qwen3.8-27B-UD3-Q4_K_M-MLX | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): size and allocation description only; card explicitly disclaims numerical parity [L24, L32] | main | |
| maccelerate/Qwen3.8-27B-UD3-Q4_K_XL-MLX | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): size and bit-width allocation only [L24,L44]; card disclaims numerical parity with the GGUF [L32] | main | |
| maccelerate/Qwen3.8-27B-UD3-Q5_K_M-MLX | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): size and bits-per-weight only; card disclaims numerical parity [L24,L32] | main | |
| maccelerate/Qwen3.8-27B-UD3-Q6_K-MLX | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): size and allocation only; card claims no numerical parity [L24,L32] | main | |
| maccelerate/Qwen3.8-27B-UD3-Q6_K_M-MLX | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): size (22.94 GB) and effective bits per weight (6.716) only; card explicitly claims no numerical parity [L24, L32] | main | |
| maccelerate/Qwen3.8-27B-UD3-Q8_K_XL-MLX | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): size and allocation only; card states no numerical parity claim with the source GGUF [L24,L32,L57] | main | |
| mahdisml/Qwen3.8-27B-Q5_0-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| malaiwah/Qwen3.8-27B-AWQ-INT4-archival-63768c10 | read in full: no quality claims | AWQ | No quality claims for the quant (card read in full): archival mirror of cyankiwi/Qwen3.8-27B-AWQ-INT4; benchmark tables [L121-283] are the upstream Qwen3.8-27B base model results, not this quant; only model size stated [L57] | main | |
| malaiwah/Qwen3.8-27B-GGUF-archival-f1bfb127 | read in full: no quality claims | GGUF | main | ||
| malaiwah/Qwen3.8-27B-MTP-NVFP4-archival-6d98dc1f | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| manubc33/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Marcin116/Qwen3.8-27B-W4A16-AutoRound-fast-embed-int4 | not yet reviewed | AutoRound | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| MariaLover/pplx-computer-qwen-3-8-27b-dflash2-20260824 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| MarxistLeninist/Qwen3.8-27B-IQ1_M-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): runtime smoke tests, speed, VRAM and tool-call pass/fail checks only; no PPL, KLD or benchmark scores [L55-94] | main | |
| Mary114514/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mattbucci/Qwen3.8-27B-AWQ | read in full: no quality claims | AWQ | main | ||
| matvei-aleksandrovich/Qwen3.8-27B-MLX-4bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Max73333/Qwen3.8-27B-GPTQ-Int4-V100 | not yet reviewed | GPTQ | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| maxidistination/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| medismera/Qwen3.8-27B-Surgical-Abliterated | not yet reviewed | other | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| medismera/Qwen3.8-cyber-RedTeam-Surgical-Abliterated | read in full: no quality claims | other | No quality claims for the quant (card read in full): native FP8 weights, no quant fidelity numbers; validation table is task scenarios with status only [L76,L89] | main | |
| mensaprodigy/Qwen3.8-27B-oQ6e-mtp-mlx | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| merve/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| meteorlad/Qwen3.8-27B-Uncensored-MLX | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Mia-AiLab/Qwen3.8-27B-DFlash2-EXL3-5.0bpw | read in full: no quality claims | EXL3 | No quality claims for the quant (card read in full): draft acceptance length and decode speed only [L57-63] | main | |
| Mia-AiLab/Qwen3.8-27B-EXL3 | read in full: no quality claims | EXL3 | No quality claims for the quant (card read in full): links to bpw branches only, no quality or size numbers [L26-45] | main | |
| michaelw9999/Qwen3.8-27B-MXFP8-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mickyba/Qwen3.8-27B-3bit-mlx | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Mike0021/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mikeinnyc/Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16 | not yet reviewed | GPTQ | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mikeyoubeach/qwen3.8-27b-uncensored-ninfer-v3 | not yet reviewed | other | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| mincon/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| mingxianderen/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mingxianderen/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| minte1431/Qwen3.8-27B-OBLITERATED | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| minte1431/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| MIRALABS/Qwen3.8-27B-W4A16-AutoRound | not yet reviewed | AutoRound | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| MisterSnrub/Qwen3.8-27B-PARO-MXFP4-v2 | read in full: no quality claims | MXFP4 | No quality claims for the quant (card read in full): only an R9700 serving measurement table with no visible numbers [L14-26]; recipe text has no quality numbers | main | |
| mistrjirka/Qwen3.8-27B-PXQ-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mixbits/Qwen3.8-27B-NVFP4-MTP-VL-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mjsxi/qwen-3.8-27b-mlx-mtp-mxfp4 | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mjsxi/qwen-3.8-27b-mlx-mtp-mxfp8 | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| MKRWW/Qwen3.8-27B-int4-AutoRound | not yet reviewed | AutoRound | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mlasli/Qwen3.8-27B-Heretic-Abliterated-MLX-8bit | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): no quant-fidelity metric (no KLD/ppl); refusal and capability tables are for the BF16 checkpoint [L53-59, L74-83]; 8-bit smoke-test rows are refusal rates only [L64-69] | main | |
| mlasli/Qwen3.8-27B-Heretic-Uncensored-IQ2_M-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| mlasli/Qwen3.8-27B-Heretic-Uncensored-IQ4_XS-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| mlasli/Qwen3.8-27B-Heretic-Uncensored-Q4_K_M-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| mlasli/Qwen3.8-27B-Heretic-Uncensored-Q5_K_M-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| mlasli/Qwen3.8-27B-Heretic-Uncensored-Q6_K-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| mlasli/Qwen3.8-27B-Heretic-Uncensored-Q8_0-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): sizes table plus a refusal-rate smoke test (behavioral direct/refuse percentages), no PPL, KLD or benchmark fidelity numbers [L42-62] | main | |
| mlx-community/Qwen3.8-27B-4bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mlx-community/Qwen3.8-27B-8bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mlx-community/Qwen3.8-27B-MTP-4bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mlx-community/Qwen3.8-27B-MTP-8bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mlx-community/Qwen3.8-27B-MTP-mxfp4 | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mlx-community/Qwen3.8-27B-MTP-mxfp8 | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mlx-community/Qwen3.8-27B-MTP-nvfp4 | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mlx-community/Qwen3.8-27B-mxfp4 | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mlx-community/Qwen3.8-27B-mxfp8 | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mlx-community/Qwen3.8-27B-nvfp4 | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| movez66/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| movez66/Qwen3.8-27B-Q5_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| movez66/Qwen3.8-27B-Q8_0-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| mradermacher/Qwen3.8-27B-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): quant size table only [L38-52]; card says sorted by size, not necessarily quality [L36] | main | |
| mradermacher/Qwen3.8-27B-i1-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): size table only with qualitative notes [L40-56] | main | |
| MrDiDi/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| mstyslavity/Qwen3.8-27B-mlx-8Bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| muhammad-taqi512/LYRA-prism | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| muxiln/Qwen3.8-27B-int4-AutoRound | not yet reviewed | AutoRound | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Muyuxiao/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| N8STORM/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| nathansutton/Qwen3.8-27B-UD-Q3_K_XL-DFlash2-MLX | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): decode speed table only [L48-53]; footprint [L27]; perplexity mentioned as selection method without numbers [L30] | main | |
| nathansutton/Qwen3.8-27B-UD-Q3_K_XL-MTP-MLX | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): size/footprint and scheme description only, no quality numbers [L22-26] | main | |
| navispace/Qwen3.8-27B-W8A16-GPTQ | read in full: no quality claims | GPTQ | main | ||
| nchapman/Qwen3.8-27B-W4A16-GPTQ | read in full: no quality claims | GPTQ | main | ||
| ndraos/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| neroued/Qwen3.8-27B-NInfer | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| nesoai-dev-org/prism-ml-Ternary-Bonsai-2-27B-gguf | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| nesoai-dev/prism-ml-Ternary-Bonsai-2-27B-gguf | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| neuralforgequantum/Qwen3.8-27B-DFlash2-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): draft-model file sizes [L40-44] and speculative-decoding acceptance length table only [L91-95]; no fidelity or quality metric | main | |
| NeuralNet-Hub/Qwen3.8-27B-Uncensored-NVFP4 | not yet reviewed | NVFP4 | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| NexVeridian/Qwen3.8-27B-5bit | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| NexVeridian/Qwen3.8-27B-6bit | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| NexVeridian/Qwen3.8-27B-8bit | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Nicolassuez/Qwen3.8-27B-W4A16 | not yet reviewed | INT4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Nicolassuez/Qwen3.8-27B-W4A16-GDN | read in full: no quality claims | INT4 | main | ||
| Nielk38/Qwen3.8-27B-MTP-GGUF-SPLIT | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| NikDowse/Qwen3.8-27B-Q6_K-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| NikiKrutan/Qwen3.8-27B-MTP-GGUF | read in full: no quality claims | GGUF | main | ||
| NilVoid/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| ninfer-5080/Qwen3.8-27B-RTX5080 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| nom666/Qwopus3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| noname222/Swift-Qwen3.8-27B-Uncensored-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Nonyabizneza/Ternary-Bonsai-2-27B-gguf | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| nova-quill/Huihui-Qwen3.8-27B-abliterated-mlx | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| nova-quill/Huihui-Qwen3.8-27B-abliterated-unsloth-mlx | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| NoxNotreve/Ternary-Bonsai-2-27B-gguf | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| npario/Qwen3.8-27B-8bit | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| npario/Qwen3.8-27B-Abliterated-MLX-6bit | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| npario/Qwen3.8-27B-MLX-5bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| npario/Qwen3.8-27B-Ridge-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| npario/Qwen3.8-27B-Uncensored-MLX | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| npario/Ternary-Bonsai-2-27B-gguf | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| nuottroisaoduoc/Bonsai-2-27B-Ternary-CRACK-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| nurdich/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| nvidia/Qwen3.8-27B-NVFP4 | read in full: no quality claims | NVFP4 | main | ||
| nvythong/Qwen3.8-27B-mlx-4Bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| nvythong/Qwen3.8-27B-mlx-6Bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| nvythong/Qwen3.8-27B-mlx-8Bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| nwtgck/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| nydemeth/Qwen3.8-27B-Uncensored-Q4_0-ROCmFP4-STRIX | not yet reviewed | NVFP4 | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Nyrud99/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| OBLITERATUS/Qwen3.8-27B-OBLITERATED | read in full: no quality claims | other | main | ||
| offmonreal/Qwen3.8-27B-MaxQuality-iMatrix-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): speed/draft-acceptance tables and file sizes only [L22-65, L85-91]; benchmark tables at L223-385 are the upstream Qwen3.8-27B model card, not the quants | main | |
| omidthekaiser/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| OnlyTextLLMs/Qwen3.8-27B-OnlyText-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): file sizes and speed tables only [L25-29, L39-46] | main | |
| OpenVINO/Qwen3.8-27B-int4-ov | not yet reviewed | INT4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| OpenVINO/Qwen3.8-27B-int8-ov | not yet reviewed | INT8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| OptimizeLLM/Qwen3.8-27B-heretic-MTP-FP8 | read in full: no quality claims | FP8 | No quality claims for the quant (card read in full): only abliteration KL (0.065 vs base on harmless prompts, Heretic ruler) and MTP acceptance/speed stated, no quant quality measurement [L29, L63-69, L106] | main | |
| orcarouter/Qwen3.8-27B-Uncensored-MLX | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| oroboros-labs/claude-opus-9-undecillion-9u5go5-sse | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| OsaurusAI/Qwen3.8-27B-JANG_4D | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| OsaurusAI/Qwen3.8-27B-JANGH2 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| OsaurusAI/Qwen3.8-27B-MXFP8 | read in full: no quality claims | MXFP8 | main | ||
| Oscilla/Qwen3.8-27B-4bit | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| oshinoWng/Qwen3.8-27B-Uncensored-HauhauCS | not yet reviewed | other | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| ovi1337/Bonsai-2-27B-CRACK-Ternary-oMLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| OzuKun/Qwen3.8-27B-FP8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| p4ik/Qwen3.8-27B-MLX-4bit | read in full: no quality claims | MLX | main | ||
| p4ik/Qwen3.8-27B-MLX-8bit | read in full: no quality claims | MLX | main | ||
| p4ik/Qwen3.8-27B-MLX-OptiQ-4bit | read in full: no quality claims | MLX | main | ||
| p4ik/Qwen3.8-27B-MLX-OptiQ-5bit | read in full: no quality claims | MLX | main | ||
| P4pps3n/Qwen3.8-27B-exl3-4.0bpw-hb8-single3090 | read in full: no quality claims | EXL3 | main | ||
| PaddyThePonderer/qwen3.8-27b-c40-uncensored | not yet reviewed | other | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| palmfuture/Qwen3.8-27B-GPTQ-Int4 | read in full: no quality claims | GPTQ | main | ||
| Paprykulus/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| parsimo2010/Qwen3.8-27B-Q4_0_ROCMFP4_STRIX-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): speed, memory and draft-acceptance tables only [L251-L437]; card states KV-cache quantization quality was not measured [L457-L460] | main | |
| PassingByPixels/Qwen3.8-27B-NVFP4 | read in full: no quality claims | NVFP4 | No quality claims for the quant (card read in full): explicitly no quality evaluation run; size and throughput only [L29-31, L201-203] | main | |
| peculiar-ragdoll/Dirk-Qwen3.8-27B-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): file size table only [L95-108]; template benchmarks at L74-76 are on a different model (Dagger/ThinkingCap-27B); ISTA quality comparison is described without numbers and attributed to ISTA [L119-123] | main | |
| peculiar-ragdoll/Dirk-Qwen3.8-27B-MLX-oQ4e | read in full: no quality claims | MLX | main | ||
| peculiar-ragdoll/Dirk-Qwen3.8-27B-MLX-oQ6e | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): tables at L68-72 are template-effect results measured on the Dagger (Qwen3.6-based) model; Dirk's own results are in-progress images [L31-39,L75]; quant table is size only [L89-92] | main | |
| perplexity-ai/pplx-computer-qwen-3-8-27b-dflash2-20260824 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| perplexity-ai/pplx-computer-qwen-3-8-27b-dflash2-gguf-20260826 | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): file sizes [L26-28] and speculative-decoding acceptance/accepted length on CPU [L59-62] only; no quant fidelity numbers | main | |
| perplexity-ai/pplx-qwen-3-8-27b-dflash2-20260819 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| pertai/Qwen3.8-27B-ET | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| pertai/qwen38-et-27b-GGUF | read in full: no quality claims | GGUF | main | ||
| petergilani/Qwen3.8-27B-MTP-8bit | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Peutlefaire/Qwen3.8-27B-NVFP4 | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| philipg11/Qwen3.8-27B-mlx-8Bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| pinvou/Qwen3.8-27B-NVFP4-Tachyon | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| pitcany/Qwen3.8-27B-FP8 | read in full: no quality claims | FP8 | No quality claims for the quant (card read in full): benchmark tables [L66-228] are the upstream Qwen3.8-27B base model results; FP8 quality only described qualitatively as nearly identical [L19] | main | |
| PixelML/Bonsai-2-27B-W4A16 | read in full: no quality claims | INT4 | main | ||
| Plaguekind/Qwen3.8-27B-FrostDFlash-MixedLow-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): DFlash2 drafter sidecard; only file sizes, bpw, speed and draft acceptance, no fidelity or accuracy numbers vs a reference [L61-70, L166-183] | main | |
| Plana-Chan/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| pottokao/Qwen3.8-27B-NVFP4-MTP-2x16GB | read in full: no quality claims | NVFP4 | main | ||
| prabanta-dev/Qwen3.8-27B-JNS | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| prakharprasad/Qwen3.8-27B-Uncensored-GGUF | read in full: no quality claims | GGUF | main | ||
| Preyazz/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| preyy2/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| prism-ml/Ternary-Bonsai-2-27B-gguf | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| prism-ml/Ternary-Bonsai-2-27B-gguf-dev | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| prithivMLmods/Qwen3.8-27B-FP8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| prithivMLmods/Qwen3.8-27B-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): only restates Qwen's published base-model benchmark scores [L31]; no quant-level quality number | main | |
| ProCreations/Ternary-Bonsai-2-27B-MTP | read in full: no quality claims | other | No quality claims for the quant (card read in full): decode tokens/sec and MTP draft acceptance only; objective and vision checks stated as historical and not rerun [L27-42] | main | |
| pugant/Qwen3.8-27B-imatrix | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| pugant/Qwen3.8-27B-MTP-Q4_0_ROCMFP4_STRIX_LEAN | read in full: no quality claims | NVFP4 | No quality claims for the quant (card read in full): speed/decode and MTP acceptance benchmarks only, no perplexity/KLD/accuracy numbers [L52-99,L131-144] | main | |
| qaqab/Qwen3.8-27B-Uncensored-Q4_K_M-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| qaqab/Qwen3.8-27B-Uncensored-Q4_K_M-Vision-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| qqwu/Qwen3.8-27B-Uncensored-IQ4-XS-MTP-16GB-VRAM-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| QQZ2026/Qwen3.8-27B-NVFP4-Q5K-no-MTP-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| QQZ2026/Qwen3.8-27B-ZeroRefusal-UD-IQ4_XS-MTP-GGUF | read in full: no quality claims | GGUF | main | ||
| qtum/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| quark75/Qwen3.8-27B-MXFP4-GGUF | read in full: no quality claims | GGUF | main | ||
| Qui-Linta13/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| quimmedes/Qwen3.8-27B-XYZ | read in full: no quality claims | other | No quality claims for the quant (card read in full): size/BPW tables only [L48-70, L178-198] | main | |
| Qwen/Qwen3.8-27B-FP8 | read in full: no quality claims | FP8 | No quality claims for the quant (card read in full): only qualitative 'nearly identical' claim [L16]; benchmark tables [L63-L216] are for the original Qwen3.8-27B, not the FP8 quant | main | |
| RadixArk/Qwen3.8-27B-NVFP4 | read in full: no quality claims | NVFP4 | main | ||
| rafw007/Qwen3.8-27B-redteam-abliterated | read in full: no quality claims | other | No quality claims for the quant (card read in full): decode/prompt speed table and qualitative behavioral test notes only [L58-62,L72-92] | main | |
| randmaru/Qwen3.8-27B-mlx-mxfp4-text-generation-only | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| rapid-mlx/Qwen3.8-27B-4bit-MTP-MLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| rAVEUK/Qwen3.8-27B-GSQ-RCO-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Ravirajabhat154/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| rawmodels/Qwen3.8-27B-antislop-semdiv-grpo-bnb-4bit | read in full: no quality claims | bitsandbytes | No quality claims for the quant (card read in full): tables are before/after fine-tuning style metrics measured on bf16 weights; card says not to quote them for this 4-bit copy [L11,L28-37] | main | |
| RedHatAI/Qwen3.8-27B-INT4 | not yet reviewed | INT4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| RedHatAI/Qwen3.8-27B-NVFP4 | read in full: no quality claims | NVFP4 | main | ||
| RESMP-DEV/Qwen3.8-27B-MLX-GPTQ-MXFP4 | read in full: no quality claims | MLX | main | ||
| Riyan200324200324/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| RKNNAI/RK1828-LLM-Qwen3.8-27B | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| robertbschuette/Qwen3.8-27B-OBLITERATED | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| RonnieOps/Qwen3.8-27B-IQ4_XS | read in full: no quality claims | other | main | ||
| ross-dev/SexyGPT-v3-Thinking-Female-gguf | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): only SFT eval loss, held-out reply rate and reward-model accuracy, no quant quality measurement [L226-235] | main | |
| RukaRat/Qwen3.8-27B-INT8-W8A8-imatrix-MTP | read in full: no quality claims | INT8 | main | ||
| RukaRat/Qwen3.8-27B-W4A16-imatrix-heretic-MTP | read in full: no quality claims | INT4 | main | ||
| RukaRat/Qwen3.8-27B-W4A16-imatrix-int8head-MTP | read in full: no quality claims | INT8 | main | ||
| Ruurd/Qwen3.8-27B-ONNX-WebGPU-INT4 | not yet reviewed | ONNX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Ryn1998/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| rzgar/Qwen3.8-27B-NVFP4-ComfyUI | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| sakamakismile/Qwen3.8-27B-MTP-NVFP4 | read in full: no quality claims | NVFP4 | No quality claims for the quant (card read in full): throughput table and size only [L25-34]; no KLD/PPL/benchmark vs bf16 | main | |
| sanbanfu/Qwen3.8-27B-ShortThink-NVFP4-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| SandLogicTechnologies/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Saviitt/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| scottlowry/Qwen3.8-27B-oQ4e-fp16-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| scottlowry/Qwen3.8-27B-oQ4e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| scottlowry/Qwen3.8-27B-oQ5e-fp16-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| scottlowry/Qwen3.8-27B-oQ5e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| scottlowry/Qwen3.8-27B-oQ6e-fp16-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| scottlowry/Qwen3.8-27B-oQ6e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| scottlowry/Qwen3.8-27B-oQ8e-fp16-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| scottlowry/Qwen3.8-27B-oQ8e-mtp | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| seabit-ai/Qwen3.8-27B-DFlash2-4bit | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| senaro/atlasv23-qwen3-8-27b-gguf | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| sensitivecontent/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| serajfast4/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| SeraphicHero/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| SergiioB/Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16 | read in full: no quality claims | GPTQ | main | ||
| SGLabs/Qwen3.8-27B-Pym-IQ2_XXS-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Sh200126/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| shawnw3i/Huihui-Qwen3.8-27B-abliterated-AWQ-MTP | not yet reviewed | AWQ | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| shawnw3i/Qwen3.8-27B-AWQ-MTP | not yet reviewed | AWQ | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| shawnw3i/Qwen3.8-27B-GPTQ-INT8-MTP | not yet reviewed | GPTQ | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| shawon/Qwen3.8-27B-mlx-4Bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| shawon/Qwen3.8-27B-mlx-8Bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| sheovsn/Ternary-Bonsai-2-27B-gguf | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Shiftedx/Qwen3.8-27B-Abliterated-MLX-MXFP4 | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Shiftedx/Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP | read in full: no quality claims | MLX | main | ||
| Shiftedx/Qwen3.8-27B-Abliterated-MLX-MXFP8 | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Shiftedx/Qwen3.8-27B-MLX-MXFP4 | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Shiftedx/Qwen3.8-27B-MLX-MXFP4-MTP | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Shiftedx/Qwen3.8-27B-MLX-MXFP8 | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| shoukewei/Qwen3.8-27B-OBLITERATED22 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| shoukewei/Qwen3.8-27B-OBLITERATED22-AGENT | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| shoukewei/Qwen3.8-27B-Uncensored-GGUF22 | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| shsunshining/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| shuijiaodawang/Qwen3.8-27B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS | not yet reviewed | other | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Simplepotat/Qwen3.8-27B-IQ5_KS-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): file/size/SHA table only [L23-L26]; quality assumed from predecessor recipe, no number [L19] | main | |
| sktthehacker/Qwen3.8-27B-Uncensored-MLX | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| slevinw/Qwen3.8-27B-RAM-13GB-GGUF | read in full: no quality claims | GGUF | main | ||
| SlickPulla/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Slomin/Potato-CODER-24GB | read in full: no quality claims | other | No quality claims for the quant (card read in full): validation (KL of cache types, LiveCodeBench-80) described without numeric values; fit/context recipe only [L189-L202, L206-L208] | main | |
| slopops/Qwen3.8-27B-int4-AutoRound-SAR | not yet reviewed | AutoRound | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| slothingaway/Qwen3.8-27B-Q8_0-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| SmartyJohnway/Qwen3.8-27B-Uncensored-MLX | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| SoAIHQ/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| sokada4/Qwen3.8-27B-GPTQ-Int4 | read in full: no quality claims | GPTQ | No quality claims for the quant (card read in full): serving speed table (TTFT/TPOT/throughput, MTP acceptance rate) only, no quality numbers [L54-61] | main | |
| soyrsoyr/Qwen3.8-27B-FP8-DYNAMIC-MTP | read in full: no quality claims | FP8 | No quality claims for the quant (card read in full): MTP speculative acceptance table only, no quality metric vs a reference [L71-L82] | main | |
| spotapovadm/Huihui-Qwen3.8-27B-abliterated-AWQ-MTP | not yet reviewed | AWQ | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| spotapovadm/Qwen3.8-27B-FP8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| srv-sngh/Qwen3.8-27B-mlx-4bit | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): only official full-model benchmark scores, explicitly not this quantization [L83-94]; size/memory otherwise | main | |
| ssfsvc/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| starskyzheng/ninfer-5080-qwen3.8-27b | read in full: no quality claims | other | No quality claims for the quant (card read in full): format-share table, MTP acceptance and performance only [L56-61, L114-118] | main | |
| steampunque/Qwen3.8-27B-MP-GGUF | read in full: no quality claims | GGUF | main | ||
| Steven10429/Qwen3.8-27B-UD-IQ4_XS-MLX-mtp | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): speed and draft-acceptance table only [L25-29]; greedy output token-identical to baseline [L23]; tensor relative error vs bf16 mtp.* only [L68] | main | |
| Sudi8dw/Ternary-Bonsai-2-27B-gguf | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| sudoingx/Ternary-Bonsai-2-27B-PTQ1_0-MTP-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): tok/s tables, draft acceptance and VRAM/size only [L71-99]; losslessness stated as byte-identical greedy output without a quality metric [L109] | main | |
| Suhyoduh/Qwen3.8-27B-FP8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| superkuh/Qwen3.8-27B-UD-IQ3_XXS-ASCII-Condensed | read in full: no quality claims | other | No quality claims for the quant (card read in full): vocabulary-pruning edit of unsloth UD-IQ3_XXS; table lists vocab/merge/row counts only, no quality measurement [L29-35] | main | |
| SuperLogic/Qwen3.8-27B-Flash-MTP-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| suryatmodulus/Qwen3.8-27B-S-experimental | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Suseezz/Qwen3.8-27B-Uncensored-IQ4-XS-MTP-16GB-VRAM-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Svenuks/Qwen3.8-27B-Uncensored-MLX | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| swdjojo/Qwen3.8-27B-MLX-oQ4e | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| swdjojo/Qwen3.8-27B-MLX-oQ5e | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| swdjojo/Qwen3.8-27B-MLX-oQ6e | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| swdjojo/Qwen3.8-27B-MLX-oQ8e | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| sweetand/qwen3.8-27b-1C | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| SwinliQ-AIn-2/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| TAKADOX/Bonsai-2-27B-Ternary-CRACK-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| TaterTotterson/Qwen3.8-27B-GGUF-Tater-NoThink | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| tchbcb/nqwen27 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| TechPrototyper/Qwen3.8-27B-DFlash2-fp8-vllm | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Ted17846578368756/qwen38-uncensored-mlx4b-mtp | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| tejones36/qwen3.8-27B-mlx-mixed-3_6 | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| TensorFold/Qwen3.8-27B-MLX-4bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| TensorFold/Qwen3.8-27B-MLX-6bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| TensorFold/Qwen3.8-27B-MLX-8bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| TensorFold/Qwen3.8-27B-oQ2 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| TensorFold/Qwen3.8-27B-oQ4 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| TensorFold/Qwen3.8-27B-oQ6 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| TensorFold/Qwen3.8-27B-oQ8 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| TensorVizion/Qwen-27B-Abliterated-MLX-2-Bit | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| terahertz/Qwen3.8-27B-8bit-MLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Terathox-Coder/Qwen3.8-27B-MTP-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| testmymodel112/Qwen3.8-27B-Uncensored-FP8 | not yet reviewed | FP8 | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| The-Sapient-Company/ghost-x1-w4a16 | not yet reviewed | INT4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| thebadking/Qwen3.8-27B-Heretic-Abliterated-Uncensored-P1M-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Thebeatdrops/Qwen3.8-27B-FP8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| TheCluster/Qwen3.8-27B-Heretic-MLX-4bit | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| TheCluster/Qwen3.8-27B-MLX-mixed-6bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| TheCluster/Qwen3.8-27B-MLX-mixed-9.4bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| thetinkerer/Qwen3.8-27B-TOS-v1 | read in full: no quality claims | other | No quality claims for the quant (card read in full): HolyC fine-tune; performance section header only, no numbers extracted [L15] | main | |
| TheWegemann/Qwen3.8-27B-LowGPU-NoMTP-IQ3XXXS-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| thoddnn/Qwen3.8-27B-8bit | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| thomasgardiner/Qwen3.8-27B-DFlash2-FP8 | read in full: no quality claims | FP8 | No quality claims for the quant (card read in full): draft-model acceptance length and decode throughput tables only [L35-40, L50-55]; target tokens are unchanged under greedy verification [L42] | main | |
| Tijani12/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| timothyckl/Qwen3.8-27B-6bit-MLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| timteh673/Qwen3.8-27B-Opus-Reasoning-Control-MLX-8bit | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): frozen local benchmark table compares control-bf16 vs abliterix-pass1-bf16 models (fine-tune/abliteration effects), not the MLX 8-bit quant itself [L38-51] | main | |
| tiyuvta/Qwen3.8-27B-NVFP4-MTP-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): decode tok/s and MTP draft acceptance rate only [L30-35,L63-79,L123-127,L224-231]; file sizes [L91-100]; no fidelity or accuracy number for any quant | main | |
| Tkayyyy702/Qwen3.8-27B-OBLITERATED | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| tokenfires/Qwen3.8-27B-MLX-6bit-MTPLX | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): speed and acceptance tables only [L46-51, L59-64]; quality checks mentioned without numbers [L53] | main | |
| ToPo-ToPo/Qwen3.8-27B-mlx-4bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| ToPo-ToPo/Qwen3.8-27B-mlx-8bit | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Toprak1yu/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Tostibrown/Qwen3.8-27B-5bit-affine-g64 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| treeish/Qwen3.8-27B-oQ2e-MTP-MLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| treeish/Qwen3.8-27B-oQ3e-MTP-MLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| treeish/Qwen3.8-27B-oQ4e-MTP-MLX | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| trevon/Qwen3.8-27B-GGUF-MTP | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| trevor000/Ternary-Bonsai-2-27B-RTX4070 | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| trklzz/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| troed/Qwen3.8-27B-ASCII-Condensed | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| trymirai/Qwen3.8-27B-L | read in full: no quality claims | other | No quality claims for the quant (card read in full): method description only, no quality numbers in card text [L49-53] | main | |
| trymirai/Qwen3.8-27B-M | read in full: no quality claims | other | No quality claims for the quant (card read in full): method description and size-tier comparability statement only, no quality numbers visible [L49-53] | main | |
| trymirai/Qwen3.8-27B-S-experimental | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| tsii42/Qwen3.8-27B-W8A8-Dynamic-Ascend-310P-vLLM | read in full: no quality claims | INT8 | No quality claims for the quant (card read in full): throughput, latency and functional smoke-test pass counts only; no fidelity or benchmark scores [L178-271] | main | |
| TTQHY/Qwen3.8-27B-8bit | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| turboderp/Qwen3.8-27B-exl3 | read in full: no quality claims | EXL3 | No quality claims for the quant (card read in full): branch links listing bits-per-weight only, no measurements [L19-37] | main | |
| TWG1412/Qwen3.8-27B-FP8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| ubergarm/Qwen3.8-27B-GGUF | read in full: no quality claims | GGUF | main | ||
| Ujwal087/Qwen3.8-27B-Q8_0-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| ukuwzi/Qwen3.8-27B-mlx-Q4_K_M | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| ulkaa/Qwen3.8-27B-AWQ-INT4 | not yet reviewed | AWQ | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| unsloth/Qwen3.8-27B-FP8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| unsloth/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| unsloth/Qwen3.8-27B-NVFP4 | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| unsloth/Qwen3.8-27B-unsloth-bnb-4bit | not yet reviewed | bitsandbytes | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| urakozz/Qwen3.8-27B-W4A16-g64-AutoRound-GPTQ | not yet reviewed | AutoRound | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| utautako/Qwen3.8-27B-NVFP4-MTP-Q8attn-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): speed and MTP draft-acceptance table only [L93-110]; accuracy only described qualitatively [L30-33] | main | |
| utrominon/Qwen3.8-27B-AWQ-BF16-INT8 | not yet reviewed | AWQ | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Vaibhavhome30/Qwen3.8-27B-FP8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| valeryepwi/Qwen3.8-27B-Q4_K_M-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| validi2007/Ternary-Bonsai-2-27B-gguf | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| vanchi34200/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| VantCdx/Qwen3.8-27B-Q4_0-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| vasiliy-mikhailov/Qwen3.8-27B-nvfp4-ru-NInfer | read in full: no quality claims | NVFP4 | main | ||
| vaultai/Qwen3.8-27B-8bit | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| vaultai/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| vcruz305/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| VIDraft/Qwen3.8-27B-FP8 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| VikramPal/Qwen3.8-27B-DynQuant-3bit | read in full: no quality claims | other | main | ||
| vincespeed/Qwen3.8-27B-No-MTP-GGUF-Dflash2 | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| vincespeed/Qwen3.8-27B-No-MTP-GGUF-Dflash2-Kquant | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): size/BPW and tensor-type tables only, no KLD/PPL/benchmark numbers [L13-18,L133-140] | main | |
| Vishva007/Qwen3.8-27B-W4A16-AutoRound | not yet reviewed | AutoRound | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Vishva007/Qwen3.8-27B-W4A16-AutoRound-GPTQ | not yet reviewed | AutoRound | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| VocaborSilentii/Qwen3.8_UnCen | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| vs4vijay/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Vtuber-plan/Huihui-Qwen3.8-27B-abliterated-NVFP4 | read in full: no quality claims | NVFP4 | main | ||
| vvsotnikov/Qwen3.8-27B-MLX-4bit | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): no quality or size numbers; only conversion recipe and checksum verification [L47-66] | main | |
| vvsotnikov/Qwen3.8-27B-MLX-8bit | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): conversion recipe and checksum verification only, no quality or size numbers [L47-66] | main | |
| vvsotnikov/Qwen3.8-27B-MTP-MLX-4bit | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): MTP drafter checkpoint; verification table lists only block_size, no quality number [L86-91] | main | |
| vvsotnikov/Qwen3.8-27B-MTP-MLX-8bit | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): MTP drafter checkpoint; no quality number, only block_size verification [L86-91] | main | |
| VykosX/Qwen3.8-27B-HyperQwen-W4A16-Fast-Prepared | not yet reviewed | INT4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Warumauchimmer/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| wayzer2321/Luna-27B-GGUF-Archive | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| webhie/Qwen3.8-27B-Q4-AutoRound-Code-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| weddle/Qwen3.8-27B-AutoRound-W4A16-G128 | not yet reviewed | AutoRound | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| Weirdo333/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| weqrzsddfwqe/Qwen3.8-27B-cybersecurity | not yet reviewed | other | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| Whytime/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| williamliao/Qwen3.8-27B-NVFP4-GGUF | read in full: no quality claims | GGUF | No quality claims for the quant (card read in full): speed/acceptance benchmarks and size only; card says no model-quality benchmark against BF16 was performed [L341] | main | |
| wodheifn/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| wuji2006/Qwen3.8-27B-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| xannie1/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| xCloudinfo/Qwen3.8-27B-Uncensored-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| xleeelx/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| xrachelburns/Huihui-Qwen3.8-27B-abliterated-GGUF | not yet reviewed | GGUF | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| XReyRobert/Qwen3.8-27B-GPTQ-Pro-FOEM-4bit-g128-ns256 | read in full: no quality claims | GPTQ | No quality claims for the quant (card read in full): recipe, serving profile and size (18.22 GiB) only; the Terminal-Bench Smoke24 scores 16/24 and 17/24 at L143-146 belong to the separate INT8-Head-Embeddings hybrid derivative, not this repo | main | |
| XReyRobert/Qwen3.8-27B-GPTQ-Pro-FOEM-4bit-g128-ns256-INT8-Head-Embeddings | read in full: no quality claims | GPTQ | main | ||
| xs7david/Qwen3.8-27B-Q2_K-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| xv0y5ncu/Qwen3.8-27B-GLQ-trellis-3inst-3bpw | read in full: no quality claims | other | main | ||
| xv0y5ncu/Qwen3.8-27B-GLQ-trellis-3inst-4bpw | read in full: no quality claims | other | main | ||
| yailabs/Qwen3.8-27B-Text-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| YCWTG/Qwen3.8-27B-NVFP4A16-GPTQ | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| ynotelbon/Qwen3.8-27B-Uncensored-MLX | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| YoozLabs/Qwen3.8-27B-lean-4bit-mlx | read in full: no quality claims | MLX | main | ||
| YourHighnessLA/L0xRE-27b-Low | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| yukimakura/Qwen3.8-27B-Q4_K_M-GGUF-2GB-split | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| yyrgzn/Qwen3.8-27B-FP8-EfficientThink-Uncensored-NInfer-v3 | read in full: no quality claims | FP8 | main | ||
| yyrgzn/Qwen3.8-27B-FP8-NInfer-v3 | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| z-lab/Qwen3.8-27B-DFlash2-GGUF | not yet reviewed | GGUF | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| z-lab/Qwen3.8-27B-PARO | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| z4bfg/Qwen3.8-27B-MLX-8bit | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): only functional case pass count (5/6) and peak-memory comparison vs BF16; no KLD/ppl/benchmark numbers [L147-150] | main | |
| zankich/Qwen3.8-27B-W8A16-FP8KV | not yet reviewed | FP8 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| zhangxinjie2026/Qwen3.8-27B-NVFP4 | not yet reviewed | NVFP4 | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| zherebetskyy/Qwen3.8-27B-4bit-mlx | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| zherebetskyy/Qwen3.8-27B-6bit-mlx | not yet reviewed | MLX | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| zherebetskyy/Qwen3.8-27B-mxfp8-mlx | read in full: no quality claims | MLX | main | ||
| zherebetskyy/Qwen3.8-27B-nvfp4-mlx | read in full: no quality claims | MLX | No quality claims for the quant (card read in full): size and target-VRAM table plus qualitative perplexity wording only, no numeric quality values [L44, L52-57] | main | |
| zipperlein/Qwen3.8-27B-GPTQ-W4A8-MTP-Ampere | not yet reviewed | GPTQ | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| zlaabsi/Qwen3.8-27B-OTQ-GGUF | read in full: no quality claims | GGUF | main | ||
| zomiailabs/Qwen3.8-27B-Uncensored-MLX | not yet reviewed | MLX | modified weights | Not yet reviewed: card has not been read in full; it may contain quality claims | main |
| zveloxy/Veloxy-Titan-32B | not yet reviewed | other | Not yet reviewed: card has not been read in full; it may contain quality claims | main | |
| zyhector/Qwen3.8-27B-TokenRush-int4g128 | read in full: no quality claims | INT4 | No quality claims for the quant (card read in full): only size (17.0 GB) and format description (asymmetric int4, group 128) [L17-26] | main |
How this was built
We listed every repository Hugging Face tags as a quantization of Qwen/Qwen3.8-27B (1220 on 2026-10-17) and saved each raw README.md at its pinned commit into a local ledger, so nothing is re-fetched. Cards that reproduce another publisher's numbers are detected by comparing claim lines across all cards and are not counted as independent measurements. Results published only as chart images were transcribed from their labels and checked against the marker positions in the image. Numbers were extracted with the line they came from, then a script confirmed that 4981 published values appear verbatim on their cited lines (0 failures). For every card that carried measured claims at the first pass (196 of 1220), a second reader re-extracted the numbers without seeing our data, and disagreements were resolved against the source text; the remaining cards were each read in full by one reader. Rows a card quotes from elsewhere are marked "quoted" and are not counted as measurements; rows measured on a different model are left out (0). File sizes without a card figure come from Hugging Face's file listing at the same commit.
Cards for this model were captured through the Hugging Face connector on 18 Oct 2026 and are pinned by the SHA-256 of their exact text, not by commit. Source links point to the current card, which may have changed since capture.
Not covered: claims made only off-site or only in images, 14 repos whose cards are gated (access denied to our account), repos the Hub search did not return, and repos created or changed after 2026-10-17. Found an error or a new quant? Open an issue or pull request; each claim lives in a plain data file with its source line.