QFF · Qwen/Qwen3.8-Flash-Next · Qwen3.8-27B

Every Hugging Face quant of Qwen3.8-Flash-Next, compared by what its publisher claims

All 453 repositories Hugging Face tags as quantizations of this model on 2026-10-17. 196 publish quality numbers about their own files; 3506 published values are copied from the card text at a pinned commit and checked against it mechanically. Nothing here was re-measured by us.

Compare KLD only within a compatible measurement set; the overview mixes methods. File size is not total VRAM usage: KV cache, runtime buffers, context length and offloading matter. How to interpret and cite QFF.

Repos
453
Publish a quality number
196
Of those, against BF16
91
Read, no quality number
257
Not yet reviewed
0
Measurements of rivals' files
152
Cards captured
2026-10-17

Read this first

  1. 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.
  2. 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.
  3. Sizes mean different things. This model carries a 51B-parameter n-gram table that some files keep on disk. A size marked "GPU weights" or "resident" leaves it out. GiB figures are converted to GB.

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.

Top-1 agreement with a Q8_0 build, measured by Maximilian228 in StrataReference: the publisher's own Q8_0. 5,153 tokens. Sizes are RAM in use while running. Their own file is highlighted.
Maximilian228 · Q8_0 (reference) 106 GiB RAM in use100%
Maximilian228 · NVFP4 + Q8_0 mix 105 GiB RAM in use95.0%
unsloth · UD-Q4_K_XL 77 GiB RAM in use94.1%
Maximilian228 · NVFP4 GPTQ 69 GiB RAM in use93.6%
nvidia · NVFP4 (converted to Strata) 69 GiB RAM in use92.9%
unsloth · UD-Q3_K_XL 58 GiB RAM in use90.7%
ISTA-DASLab · IQ3_S 52 GiB RAM in use90.0%
ISTA-DASLab · IQ3_XXS 45 GiB RAM in use87.3%
ISTA-DASLab · IQ2_XS 38 GiB RAM in use85.0%
ISTA-DASLab · Q2_0 37 GiB RAM in use84.1%
Mean KLD against unsloth's Q8_0, measured by agentionai (Gyro card)Held-out technical writing and code, 2,048 context, 60 chunks. Sizes are GPU-resident weights; the n-gram table is excluded. Lower is better. Their own files are highlighted.
agentionai · AP-Q4_K_XL 67.38 GiB GPU weights0.107
agentionai · Gyro-L (TQ3_0) 50.21 GiB GPU weights0.146
ggml-org · IQ4_NL 68.21 GiB GPU weights0.200
AtomicChat · AD-4.27bpw Q4_K_M 52.26 GiB GPU weights0.228
agentionai · Gyro-M (TQ2_0) 34.98 GiB GPU weights0.291
ISTA-DASLab · IQ3_XXS 43.8 GiB GPU weights0.311
unsloth · UD-Q2_K_XL 46.63 GiB GPU weights0.334
agentionai · Gyro-S (TQ1_0) 27.64 GiB GPU weights0.401
ISTA-DASLab · IQ2_XS 36.52 GiB GPU weights0.420
Mean KLD on held-out text, measured by agentionai (AP card)2,048 context. The card refers to 'the same source' without naming it. Sizes are download size. Lower is better.
agentionai · AP-Q5_K_M 103.57 GiB download0.0771 ± 0.0007
unsloth · UD-Q4_K_XL 103.69 GiB download0.0778
agentionai · AP-Q5_K_XL 112.51 GiB download0.0839 ± 0.0008
agentionai · AP-Q4_K_XL 94.2 GiB download0.0886 ± 0.0008
agentionai · AP-Q4_K_M 88.04 GiB download0.1243 ± 0.0011
agentionai · AP-IQ4_XS 84.24 GiB download0.1584 ± 0.0014
agentionai · AP-IQ3_XXS 80.72 GiB download0.1844 ± 0.0016
agentionai · AP-IQ2_S 76.03 GiB download0.2657 ± 0.0022
Top-1 agreement with BF16, measured by ivanfioravantiOwn logprob harness, 85 cases of 24 tokens. Small sample.
ggml-org · Q8_0 98.7%
ivanfioravanti · Q4KImatrixExperts-MXFP4Down-...-MTP 74.9 GB96.6%
ivanfioravanti · previous Q4_0-routed release 96.5%

Same file, different numbers

Where two cards report on the same file, the figures rarely agree. The Setup column lists what each card states about its method; where the stated setups differ, that may explain the gap, but the cause of any particular difference is not established, and the two cards may not have tested the same file revision.

FileMeasured byClaimSetupSource
unsloth UD-Q4_K_XLunsloth (own card)KLD 0.046893reference not statedunsloth.ai docs
Maximilian228top-1 94.1%vs Maximilian228's Q8_0, inside StrataL35 @ 379ff0b
agentionaiKLD 0.0778; top-1 87.21 %held-out text, reference not named on the cardL71 @ 89f38ee
unsloth UD-Q2_K_XLunsloth (own card)KLD 0.224607reference not statedunsloth.ai docs
agentionaiKLD 0.334; top-1 76.61%vs unsloth Q8_0, held-outL62 @ aa0a0c3
AtomicChat AD-4.27bpwAtomicChat (own card)KLD 0.0842; top-1 89.49%vs BF16, 87 chunks at 4,096 contextL84 @ 1c1f33a
agentionaiKLD 0.228; top-1 80.58%vs unsloth Q8_0, 2,048 contextL63 @ aa0a0c3
agentionai AP-Q4_K_XLagentionai (own card)KLD 0.0886 ± 0.0008; top-1 86.14 ± 0.14 %held-out text, reference not named on the cardL49 @ 89f38ee
agentionaiKLD 0.107; top-1 85.52%same publisher, other card: stated reference unsloth Q8_0 (the AP card does not name its reference)L65 @ aa0a0c3
ddalcu mixed-4-8bitddalcu (own card)top-1 87.9%own cardL23 @ 788c07a
ddalcu (own card)top-1 89.1%same publisher's 3.3bpw cardL55 @ 4fabeb5

The ledger

RepositoryVariantFormatSize GBSize isMean KLDTop-1PPLPPL change
vs the card's baseline		Other metrics (as published)ReferenceSource line
30cmtonyfaker/REAP-288-MLX-Serve-4bit-MTP modified weightsREAP-288 4-bit MTPMLX——————no quality claim@ b8a6c8d
6block/GGUF IQ4_XSGGUF127.6download——1.8378 ± 0.05554——vs BF16L40 @ 2d83ab7
6block/GGUF IQ3_XXSGGUF115.3download——2.0528 ± 0.06396——vs BF16L41 @ 2d83ab7
6block/GGUF IQ2_XS-layeredGGUF89.66download——2.7833 ± 0.10058——vs BF16L42 @ 2d83ab7
6block/GGUF IQ1_M-layeredGGUF82.03download——4.4105 ± 0.19322——vs BF16L43 @ 2d83ab7
AesSedai/GGUF Q5_K_MGGUF164.4download0.020218 ± 0.000145—4.596526 ± 0.027937+0.1270%—reference not statedL15 @ d933da2
AesSedai/GGUF Q4_K_MGGUF141.7download0.034206 ± 0.000237—4.613527 ± 0.028075+0.4973%—reference not statedL16 @ d933da2
AesSedai/GGUF Q5_K_M (Q4_0 PLE)GGUF138.8download0.030943 ± 0.000212—4.620130 ± 0.028097+0.6412%—reference not statedL23 @ d933da2
AesSedai/GGUF IQ4_XSGGUF123.5download0.076158 ± 0.000486—4.739310 ± 0.029099+3.2373%—reference not statedL17 @ d933da2
AesSedai/GGUF Q4_K_M (Q4_0 PLE)GGUF116.1download0.043475 ± 0.000297—4.629426 ± 0.028160+0.8437%—reference not statedL24 @ d933da2
AesSedai/GGUF IQ3_SGGUF113.8download0.160378 ± 0.000915—4.863539 ± 0.030050+5.9434%—reference not statedL18 @ d933da2
AesSedai/GGUF IQ2_SGGUF111.3download0.205353 ± 0.001121—5.043883 ± 0.031497+9.8718%—reference not statedL19 @ d933da2
AesSedai/GGUF IQ4_XS (Q4_0 PLE)GGUF97.86download0.084406 ± 0.000531—4.759109 ± 0.029230+3.6686%—reference not statedL25 @ d933da2
AesSedai/GGUF IQ3_S (Q4_0 PLE)GGUF88.19download0.167198 ± 0.000940—4.889509 ± 0.030205+6.5091%—reference not statedL26 @ d933da2
AesSedai/GGUF IQ2_S (Q4_0 PLE)GGUF85.67download0.211550 ± 0.001142—5.062449 ± 0.031623+10.2763%—reference not statedL27 @ d933da2
AetherKelly00/GGUF not statedGGUF——————no quality claim@ b5ae04a
agentionai/AP-GGUF AP-Q5_K_XLGGUF120.8download0.0839 ± 0.000886.63 ± 0.14 %———reference unclearL47 @ 89f38ee
agentionai/AP-GGUF AP-Q5_K_MGGUF111.2download0.0771 ± 0.000786.95 ± 0.14 %——kld:code 0.0181; top1:code 97.69 %reference unclearL48 @ 89f38ee
agentionai/AP-GGUF AP-Q4_K_XLGGUF101.2download0.0886 ± 0.000886.14 ± 0.14 %——kld:code 0.0206; top1:code 97.38 %reference unclearL49 @ 89f38ee
agentionai/AP-GGUF AP-Q4_K_MGGUF94.53download0.1243 ± 0.001184.22 ± 0.15 %———reference unclearL50 @ 89f38ee
agentionai/AP-GGUF AP-IQ4_XSGGUF90.45download0.1584 ± 0.001482.89 ± 0.15 %———reference unclearL51 @ 89f38ee
agentionai/AP-GGUF AP-IQ3_XXSGGUF86.67download0.1844 ± 0.001681.85 ± 0.16 %———reference unclearL52 @ 89f38ee
agentionai/AP-GGUF AP-IQ2_SGGUF81.64download0.2657 ± 0.002279.14 ± 0.16 %———reference unclearL53 @ 89f38ee
agentionai/Gyro-GGUF Gyro-L (TQ3_0)GGUF53.91GPU weights0.14683.87%——kld:code 0.027vs a Q8_0L59 @ aa0a0c3
agentionai/Gyro-GGUF Gyro-M (TQ2_0)GGUF37.56GPU weights0.29178.75%——kld:code 0.066vs a Q8_0L58 @ aa0a0c3
agentionai/Gyro-GGUF Gyro-S (TQ1_0)GGUF29.68GPU weights0.40175.29%——kld:code 0.109vs a Q8_0L57 @ aa0a0c3
agentionai/MTP-Q8_0-GGUF Qwen3.8-Flash-Next-MTP-Q8_0.ggufGGUF (MTP draft head only)4.1download—————no quality claimL20 @ 640aed4
agentionai/MTP-ROCmFP4-FAST-GGUF MTP-ROCmFP4-FAST draft headGGUF (ROCmFP4, MTP draft head only)2.5download—————no quality claimL23 @ 5a2cf56
agentionai/ROCmFP4-FAST-imatrix-GGUF ROCmFP4-FASTGGUF93.48download——4.1062 +/- 0.02329+2.48%—vs BF16L161 @ d37a14a
agnosticeng/4bit 4-bit gs32MLX-style (lisa runtime)——————no quality claim@ 97a3bd1
ahadadada/Ahadadada-FP8 FP8FP8185.5download—————no quality claimL16 @ 8a311cf
airagrp/mixed-6-8-ple8-mlx-serve mixed 6/8, PLE 8-bitMLX (mlx-serve pack)105.2trunk only—————no quality claimL114 @ b909fe3
airagrp/mixed-iQ5-Q8-ple16-mlx-serve mixed-iQ5-Q8-ple16MLX202.0download0.072 kld92.1%———vs BF16L116 @ af5fc98
aixiaoma/W4A16 W4A16, TP8+EPcompressed-tensors (W4A16)168.0download————total 48/50; arith 20/20; json 10/10; nth 9/10; sort 9/10other or no referenceL94 @ 75234a1
aixiaoma/W4A16 W4A16, TP4+EPcompressed-tensors (W4A16)—————total 47/50; arith 20/20; json 9/10; nth 9/10; sort 9/10other or no referenceL95 @ 75234a1
Alexsi98/Huihui-abliterated-GGUF modified weightsUD-Q4_K_XLGGUF——————no quality claim@ 7711281
amd/EXTENSOR ROCmFP4 EXTENSOREXTENSOR GGUF (not llama.cpp compatible)176.7download—————no quality claimL26 @ c55e64b
amd/Quark-MXFP4 MXFP4MXFP4—————GSM8K (5-shot) 96.5; recovery_pct 99.7%vs BF16L86 @ b2091c3
amd/Quark-MXFP4-AttnFP8-PLEFP8 Quark-MXFP4-AttnFP8-PLEFP8MXFP4+FP8—————GSM8K (5-shot) 96.21; recovery_pct 99.3%vs BF16L94 @ bd6e49d
amd/Quark-MXFP4-PLEFP8 MXFP4-PLEFP8MXFP4—————GSM8K (5-shot) 96.44; recovery_pct 99.5%vs BF16L93 @ 1ad36fa
aniolekx/E9-MLX-8bit E9 (8-bit g64, n-gram 8-bit g32)MLX195.0download————ppl:technical 6.072; top1:technical 64.57%; nll_delta_vs_predecessor:technical 0.110 nats/token better (t = 16.5); nll_delta_vs_predecessor:prose −0.00096, t = −0.66other or no referenceL103 @ cfd4a00
AnonimousA/REAP-256-duo-GGUF modified weightsUD-Q3_K_XL-reap256GGUF61.94download————battery:medium 12–13 / 13; battery:low 12 / 13; HumanEval 89.6%; routing_mass_retained 89.29%other or no referenceL70 @ c806923
AnonimousA/REAP-320-GGUF modified weightsUD-Q3_K_XL-lineage REAP-320 (main)GGUF68.93download————6-task battery (3 passes) 13.0 / 13.0 / 12.0; HumanEval-164 pass@1 96.3%; rare_name_probe 9/10; fabrication_rate 25%other or no referenceL63 @ 4bc31ec
AnonimousA/REAP-320-GGUF modified weightsQ2 (UD-Q2_K_XL-lineage) REAP-320GGUF61.53download————HumanEval-164 pass@1 95.1%; rare_name_probe 9/10; fabrication_rate 25%; HumanEval-164 pass@1 with MTP 94.5%other or no referenceL48 @ 4bc31ec
antirez/gguf Q4GGUF177.3download—————no quality claimL23 @ d600fe1
antirez/gguf Q2GGUF147.2download—————no quality claimL22 @ d600fe1
apetersson/Abliterated modified weightsStock BF16 n-gram table shard (shard 2)GGUF102.4download—————other or no referenceL55 @ 89bb1c9
apetersson/Abliterated modified weightsAbliterated Q5_K_M (baked trunk)GGUF—————IFEval 7/8; MATH-500 7/8; HumanEval+/MBPP+ 10/12; MMLU-Pro slice 6/8; proj_retained:Q5_K 0.97–1.00; proj_retained:Q8_0 0.99–1.01; proj_retained:Q5_1 0.86; proj_retained:IQ4_NL 0.34other or no referenceL38 @ 89bb1c9
apetersson/GGUF Q5_K_M trunkGGUF81.86trunk only—————no quality claimL57 @ 9af845f
apetersson/GGUF UD-IQ4_XS trunkGGUF64.88trunk only—————no quality claimL58 @ 9af845f
Ar3ntiz/qwen3.8_flash_next_side8 SUC1 side8SUC196.32download0.0640.913 top14.704—Total 97/120; GSM8K 17; MATH 15; MMLU-Pro 6; ARC-C 20; BBH 19; Long context 20vs BF16L49 @ ba49b29
argyelan/GGUF IQ4XS-NGQ5GGUF98.00download————HumanEval 97.0 %; HumanEval (passed/total) 159/164; GSM8K 96.0 %; GSM8K (correct/total) 192/200other or no referenceL56 @ 22e8725
arnomatic/Qwen38-AWQ-g32-ple8 AWQ-g32 + PLE8AWQ (moe_wna16) + W8A16 PLE131.0download————ple_corr_vs_ref_w8a16 0.9996; ple_max_numeric_diff 2.29e-04other or no referenceL30 @ d40f5e9
ArtomYuan/ROCmFPX Q8_0_ROCMFPX_AGENTGGUF (ROCmFPX extension types)186.4download—————no quality claimL82 @ e194a61
ArtomYuan/ROCmFPX Q8_0_ROCMFPXGGUF (ROCmFPX extension types)182.7download—————no quality claimL81 @ e194a61
ArtomYuan/ROCmFPX Q6_0_ROCMFPX_AGENTGGUF (ROCmFPX extension types)169.7download—————no quality claimL80 @ e194a61
ArtomYuan/ROCmFPX Q6_0_ROCMFPXGGUF (ROCmFPX extension types)145.7download—————no quality claimL79 @ e194a61
ArtomYuan/ROCmFPX Q4_0_ROCMFP4_STRIXGGUF (ROCmFPX extension types)121.7download—————no quality claimL78 @ e194a61
ArtomYuan/ROCmFPX Q4_0_ROCMFP4_FAST_COHERENTGGUF (ROCmFPX extension types)121.7download—————no quality claimL77 @ e194a61
ArtomYuan/ROCmFPX Q4_0_ROCMFP4_FASTGGUF (ROCmFPX extension types)94.26download—————no quality claimL76 @ e194a61
ashbash/MTP-Drafter-GGUF mtp-drafter-Q8_0GGUF (MTP draft head)2.8download—————no quality claimL32 @ c82053e
ashbash/MTP-Drafter-GGUF mtp-drafter-Q4_K_LGGUF (MTP draft head)1.9download—————no quality claimL33 @ c82053e
ashuaria/GGUF-FineTuning not listed on cardGGUF——————no quality claim@ 147b9cc
AtomicChat/GGUF AD-5.00bpw-Q5_K_M-M64GGUF110.5download0.083789.55%—1.026—vs BF16L85 @ 1c1f33a
AtomicChat/GGUF AD-4.27bpw-Q4_K_M-M64GGUF94.50download0.084289.49%—1.026—vs BF16L84 @ 1c1f33a
AtomicChat/GGUF AD-3.84bpw-IQ4_XS-M64GGUF84.90download0.227782.68%—1.102—vs BF16L83 @ 1c1f33a
auggie246/Swift-1.5-W4A16-FP8PLE modified weightsSwift 1.5 W4A16 + FP8 PLE + MTP3AutoRound W4A16 (auto_gptq) + FP8 PLE120.8download—————no quality claimL57 @ aecff07
auto7x/OrcaRouter-Uncensored-ModelOpt-NVFP4-GGUF modified weightsModelOpt NVFP4NVFP4 (ModelOpt safetensors)135.3download—————no quality claimL51 @ 3747631
AutomatosX/AX-MLX-AXQ-MXFP4-MTP AXQ MXFP4 budget class (incl. MTP)MLX132.3download—————no quality claimL48 @ 4fdb4a6
AutomatosX/AX-MLX-AXQ-MXFP8-MTP AXQ-MXFP8MLX (AXQuant)189.2download—————no quality claimL62 @ 48b9aa5
axiomofmind/W4A16-NVFP4 W4A16 NVFP4NVFP4 (ModelOpt, W4A16)——————no quality claim@ 41ef3b0
AxionML/NVFP4 NVFP4NVFP4135.0download————GSM8K 97.27; AIME 2026 pass@1 98.75vs BF16L48 @ 21d023c
Baekpica/GGUF BF16GGUF (custom qwen4exp schema, ds4)360.0download—————no quality claimL45 @ 5dfe6f0
Baekpica/GGUF Q8_0GGUF (custom qwen4exp schema, ds4)192.2download—————no quality claimL58 @ 5dfe6f0
Baekpica/Mixed-Quant-GGUF MQ-NG5-Q2K-Q4-SPLIT-E4GGUF (custom qwen4exp schema, ds4)98.50download—————no quality claimL52 @ 151ad0c
Baekpica/Mixed-Quant-GGUF MQ-NG5-Q4-SPLIT-E5 (projected, not built)GGUF (custom qwen4exp schema, ds4)93.85download—————no quality claimL44 @ 151ad0c
Baekpica/Mixed-Quant-SSD-PLE-GGUF PLE sidecar BF16 (shared, SSD-resident)GGUF (custom qwen4exp schema, ds4 SSD-PLE)102.4download—————other or no referenceL67 @ f92da01
Baekpica/Mixed-Quant-SSD-PLE-GGUF MQ-Q6-SSD-PLE-BF16GGUF (custom qwen4exp schema, ds4 SSD-PLE)97.67download————ppl:Long-form essay 8.882991; nll:Long-form essay 2.184138; ppl:Repetitive structured security fixture 1.071435; nll:Repetitive structured security fixture 0.068999; ppl:Equal-token aggregate 3.085053; nll:Equal-token aggregate 1.126569other or no referenceL1119 @ f92da01
Baekpica/Mixed-Quant-SSD-PLE-GGUF MQ-Q5-SSD-PLE-BF16GGUF (custom qwen4exp schema, ds4 SSD-PLE)83.28download————ppl:Long-form essay 8.502567; nll:Long-form essay 2.140368; nll_delta_vs_q6:Long-form essay -0.043770 (-2.00%); ppl:Repetitive structured security fixture 1.084437; nll:Repetitive structured security fixture 0.081061; nll_delta_vs_q6:Repetitive structured security fixture +0.012062 (+17.48%); ppl:Equal-token aggregate 3.036527; nll:Equal-token aggregate 1.110714; nll_delta_vs_q6:Equal-token aggregate -0.015854 (-1.41%)other or no referenceL1119 @ f92da01
Baekpica/Mixed-Quant-SSD-PLE-GGUF PLE sidecar FP8 E4M3FN (optional)GGUF (custom qwen4exp schema, ds4 SSD-PLE)51.20download—————other or no referenceL68 @ f92da01
bartowski/GGUF bf16GGUF354.0download—————no quality claimL44 @ db3a185
bartowski/GGUF Q8_0GGUF188.3download—————no quality claimL45 @ db3a185
bartowski/GGUF Q6_KGGUF168.1download—————no quality claimL46 @ db3a185
bartowski/GGUF Q5_K_LGGUF151.1download—————no quality claimL47 @ db3a185
bartowski/GGUF Q4_K_LGGUF139.3download—————no quality claimL48 @ db3a185
bartowski/GGUF Q5_K_MGGUF134.7download—————no quality claimL49 @ db3a185
bartowski/GGUF Q5_K_SGGUF128.1download—————no quality claimL50 @ db3a185
bartowski/GGUF Q4_K_MGGUF119.6download—————no quality claimL51 @ db3a185
bartowski/GGUF Q3_K_XLGGUF119.3download—————no quality claimL52 @ db3a185
bartowski/GGUF Q4_K_SGGUF113.1download—————no quality claimL53 @ db3a185
bartowski/GGUF Q4_1GGUF111.3download—————no quality claimL54 @ db3a185
bartowski/GGUF Q2_K_LGGUF107.2download—————no quality claimL55 @ db3a185
bartowski/GGUF Q4_0GGUF100.6download—————no quality claimL56 @ db3a185
bartowski/GGUF IQ4_NLGGUF100.3download—————no quality claimL57 @ db3a185
bartowski/GGUF IQ4_XSGGUF97.68download—————no quality claimL58 @ db3a185
bartowski/GGUF IQ3_MGGUF93.23download—————no quality claimL59 @ db3a185
bartowski/GGUF Q3_K_LGGUF93.19download—————no quality claimL60 @ db3a185
bartowski/GGUF Q3_K_MGGUF91.96download—————no quality claimL61 @ db3a185
bartowski/GGUF IQ3_XSGGUF91.95download—————no quality claimL62 @ db3a185
bartowski/GGUF Q3_K_SGGUF89.44download—————no quality claimL63 @ db3a185
bartowski/GGUF IQ3_XXSGGUF88.02download—————no quality claimL64 @ db3a185
bartowski/GGUF Q2_KGGUF80.93download—————no quality claimL65 @ db3a185
bartowski/GGUF IQ2_MGGUF80.37download—————no quality claimL66 @ db3a185
bartowski/GGUF IQ2_SGGUF77.83download—————no quality claimL67 @ db3a185
bartowski/GGUF IQ2_XSGGUF77.74download—————no quality claimL68 @ db3a185
bartowski/GGUF IQ2_XXSGGUF75.19download—————no quality claimL69 @ db3a185
bartowski/GGUF IQ1_MGGUF72.01download—————no quality claimL70 @ db3a185
bartowski/GGUF IQ1_SGGUF70.10download—————no quality claimL71 @ db3a185
batiai/GGUF Q6_K (Q5_1 PLE)GGUF151.6download————korean_hangul_pct 70.5%other or no referenceL75 @ 1d0ec19
batiai/GGUF Q4_K_MGGUF119.2download—————other or no referenceL74 @ 1d0ec19
batiai/GGUF IQ4_XSGGUF97.50download————functional_checks 7/7; korean_hangul_pct 70.4%; korean_longform_hangul_pct 73.1%other or no referenceL73 @ 1d0ec19
batiai/GGUF IQ3_XXSGGUF85.04download————korean_hangul_pct 70.0%other or no referenceL72 @ 1d0ec19
beamster/Sushi-2.6bpw Sushi-2.6bpw (4-bit n-gram table)EXL3 (sushi pack)79.46download0.1355 kld89.08%——kld:bf16_ngram_table 0.1303; LLMProbe Hard 50 88.0%vs BF16L53 @ 0658b15
beamster/Sushi-2bpw Sushi-2bpwEXL3 (sushi pack)37.55GPU weights0.208085.94%——LLMProbe Hard 50 84.0%vs BF16L57 @ cb6c85e
beamster/Sushi-3bpw Sushi-3bpw (4-bit n-gram table)EXL3 (sushi pack)52.97GPU weights0.103690.31%——LLMProbe Hard 50 82.0%vs BF16L52 @ 76bd8f5
beamster/Sushi-4bpw Sushi-4bpw (bf16 n-gram table)EXL3 (sushi pack)170.7download0.0592 kld92.89%——kld:4bit_ngram_table 0.0654; LLMProbe Hard 50 90.0%vs BF16L52 @ f25c0c4
beginning-ai/NVFP4 NVFP4 mixed (beginning-ai)NVFP4 + FP8 (ModelOpt mixed)128.8download————DeepSWE v1.1 (30-task subset) 19/30 = 63.3%; partial:new-feature tests passed 95.7%; partial:existing tests passed 97.7%; conversation_replay_wellformed 120 of 120vs BF16L92 @ 9c8cb34
Beinsezii/GGUF-HALO HALO GGUFGGUF——————no quality claim@ e1c7270
BenMonkey11/Huihui-abliterated-v2-GGUF modified weightsUD-Q4_K_XLGGUF——————no quality claim@ 85a485c
bitlamas/GGUF Q4_K_XL-IQ4NLGGUF102.5download—94.40 %2.9092 ± 0.0362—kld_vs_UD-Q4_K_XL 0.0304; down_err_vs_bf16 0.00490other or no referenceL49 @ 408a8e6
BlivionIaG/EXL3-3bpw-3inst EXL3 3inst (experts K=3, dense K=6)EXL3——————no quality claim@ ceaf21e
BlivionIaG/EXL3-3bpw-3inst-uncalibrated EXL3 3inst ~3bpw (uncalibrated)EXL3——————no quality claim@ 5c7f0eb
BoldingBuilds/Reasoning-Termination-Fix-GGUF modified weightsUD-IQ4_XS Reasoning-Termination-FixGGUF94.00download————MATH-500 (200) correct 168; MATH-500 (200) empty final response 0; HumanEval pass@1 passed 155; HumanEval empty final response 0; StrongREJECT refusal 143 (95.3%); SimpleSafetyTests refusal 84; XSTest safe full compliance 92other or no referenceL32 @ bb56765
boxrunner/IQ3_S-GGUF IQ3_S-expert (map C + IQ3_S gate/up)GGUF45.95download—————no quality claimL32 @ ffd258d
boxrunner/mixC-GGUF oQ2.0-mixC part 1GGUF47.62download—————no quality claimL30 @ 028efbb
boxrunner/mixC-GGUF oQ2.0-mixC part 2 (n-gram table)GGUF28.80download—————no quality claimL31 @ 028efbb
boxrunner/mixC-GGUF mtp-head-IQ4_NL.gguf (optional)GGUF2.4download—————no quality claimL33 @ 028efbb
brianbellx/Huihui-abliterated-GGUF-copy modified weightsUD-Q4_K_XL (per example command)GGUF——————no quality claim@ 0dec8ab
btbtyler09/GPTQ-4bit GPTQ 4-bitGPTQ188.0download0.07191.3%3.1386+0.58%kld_p99 0.89; top5_overlap 84.6%; ppl:wiki4k 2.2822; ppl_delta_pct:wiki4k −0.43%; mean_kld:wiki4k 0.093; top1:wiki4k 91.7%; ppl:code 2.1796; ppl_delta_pct:code −0.02%; kld:code 0.027; top1:code 95.8%; ppl:vllm_served 3.1362vs BF16L78 @ 9ae8d0e
butterf1ying/Uncensored-REAP288-MTP-MLX-4bit modified weightsUncensored REAP288 MLX 4-bitMLX (mlx-serve)76.96download—————no quality claimL76 @ 7fef629
cafonez/HC-Q8 HC-Q8GGUF (ROCmFPX types)92.35download—————no quality claimL21 @ 986da0f
carsenk/ROGUE-NVFP4 modified weightsROGUE NVFP4NVFP4 (ModelOpt)111.9download—————no quality claimL50 @ dbd1690
cc1244578151/GSQ-RCO-GGUF GSQ-RCO IQ3_SGGUF83.60download—————no quality claimL127 @ c14ca5b
cc1244578151/GSQ-RCO-GGUF GSQ-RCO IQ3_XXSGGUF75.80download—————no quality claimL126 @ c14ca5b
cc1244578151/GSQ-RCO-GGUF GSQ-RCO IQ2_XSGGUF68.00download—————no quality claimL125 @ c14ca5b
cc1244578151/GSQ-RCO-GGUF GSQ-RCO Q2_0GGUF66.40download—————no quality claimL124 @ c14ca5b
cc1244578151/myqwen3.8flashgsqrco GSQ-RCO IQ3_SGGUF83.60download————AIME25 100.00; GPQA-Diamond 92.93; LiveCodeBench v6 86.86; task_avg 93.26vs BF16L127 @ a6341db
cc1244578151/myqwen3.8flashgsqrco GSQ-RCO IQ3_XXSGGUF75.80download————zs_avg 77.23; recovery_pct 100.4%; AIME25 100.00; GPQA-Diamond 91.41; LiveCodeBench v6 86.29; task_avg 92.57vs BF16L126 @ a6341db
cc1244578151/myqwen3.8flashgsqrco GSQ-RCO IQ2_XSGGUF68.00download————zs_avg 77.16; recovery_pct 100.3%; AIME25 96.67; GPQA-Diamond 87.37; LiveCodeBench v6 83.43; task_avg 89.16vs BF16L125 @ a6341db
cc1244578151/myqwen3.8flashgsqrco GSQ-RCO Q2_0GGUF66.40download————zs_avg 78.00; recovery_pct 101.4%; AIME25 96.67; GPQA-Diamond 89.39; LiveCodeBench v6 81.14; task_avg 89.07vs BF16L124 @ a6341db
chatnerdie/EXL3-Sushi 5bpw-shEXL3 (sushi pack)83.54GPU weights————kld:yarn 0.004843; top1:yarn 98.4%; kld:native 0.002124; top1:native 99.0%; kld_ratio_vs_sushi4bpw:yarn 0.79 (0.72–0.86); kld_ratio_vs_sushi4bpw:native 0.57 (0.46–0.68); tool_calling 28 / 28vs BF16L63 @ 5329132
chatnerdie/EXL3-Sushi 4bpw-shEXL3 (sushi pack)68.40GPU weights————kld:yarn 0.005780; top1:yarn 98.5%; kld:native 0.003525; top1:native 98.4%; kld_ratio_vs_sushi4bpw:yarn 0.94 (0.88–1.00); kld_ratio_vs_sushi4bpw:native 0.94 (0.72–1.15); tool_calling 24 / 28vs BF16L62 @ 5329132
cmh/5.33bpw-PLEQ8_0 5.33bpw PLEQ8_0GGUF117.9download—————no quality claimL19 @ 3b07e4d
CMSManhattan/JiRackDeltaNet_180b BF16 (JiRack export)BF16 safetensors (JiRack port); GGUF planned——————no quality claim@ 82d1140
CMSManhattan/JiRackDeltaNet_180b Q8_0 (planned)BF16 safetensors (JiRack port); GGUF planned——————no quality claim@ 82d1140
CMSManhattan/JiRackDeltaNet_180b Q4_K_M (planned)BF16 safetensors (JiRack port); GGUF planned——————no quality claim@ 82d1140
CMSManhattan/JiRackDeltaNet_180b Q3_K_M (planned)BF16 safetensors (JiRack port); GGUF planned——————no quality claim@ 82d1140
CMSManhattan/JiRackDeltaNet_180b Q2_K (planned)BF16 safetensors (JiRack port); GGUF planned——————no quality claim@ 82d1140
CMSManhattan/JiRackDeltaNet_180b TQ2_0 (planned, after QAT)BF16 safetensors (JiRack port); GGUF planned——————no quality claim@ 82d1140
codelight-ai/NVFP4-PLE-INT3 NVFP4 + INT3 PLENVFP4 experts + INT3 PLE (compressed-tensors style)99.30download————ple_rel_err 0.1854; ple_cos 0.982685; ple_size_gb 20.5 GBother or no referenceL64 @ 59ecf4e
cyankiwi/AWQ-INT4 AWQ-INT4AWQ——————no quality claim@ d39638a
cygnal/Heretic2-IQ4XS-NGQ4-GGUF modified weightsHeretic2 IQ4_XS NGQ4GGUF92.00download————HumanEval+ base pass@1 84.1%; HumanEval+ plus pass@1 79.3%; abliteration_kl_vs_original 0.0818other or no referenceL85 @ b7a350f
Cyronius/131B-A6B-GGUF modified weightsqwen38-keep1-Q4_0GGUF75.40download—————other or no referenceL37 @ 2b7b5c7
Cyronius/131B-A6B-GGUF modified weightsqwen38-keep1-Q3KXLGGUF64.80download——4.66—tools 1.000; GSM8K-15 1.000; MMLU-30 .833other or no referenceL50 @ 2b7b5c7
d9beuD/oQ2-mtp oQ2MLX64.00download—————no quality claimL27 @ 120564d
d9beuD/oQ2.5-mtp oQ2.5MLX71.00download—————no quality claimL27 @ 9588892
d9beuD/oQ2.5e-mtp oQ2.5eMLX71.00download—————no quality claimL31 @ c9fd3c3
d9beuD/oQ2.7-mtp oQ2.7MLX74.00download—————no quality claimL27 @ 23bd805
d9beuD/oQ2.7e-mtp oQ2.7eMLX74.00download—————no quality claimL31 @ 77f5e63
d9beuD/oQ2e-mtp oQ2eMLX64.00download—————no quality claimL31 @ 7680df8
d9beuD/oQ3-mtp oQ3MLX84.00download—————no quality claimL27 @ cf5fd1a
d9beuD/oQ3.5-mtp oQ3.5MLX90.00download—————no quality claimL27 @ c6d197d
d9beuD/oQ3.5e-mtp oQ3.5eMLX90.00download—————no quality claimL31 @ fc2a7d1
d9beuD/oQ3e-mtp oQ3eMLX84.00download—————no quality claimL31 @ 24e392e
d9beuD/oQ4-mtp oQ4MLX106.0download—————no quality claimL27 @ 12e38a2
d9beuD/oQ4e-mtp oQ4eMLX106.0download—————no quality claimL31 @ 0b047d1
d9beuD/oQ5-mtp oQ5MLX129.0download—————no quality claimL27 @ da3149c
d9beuD/oQ5e-mtp oQ5eMLX129.0download—————no quality claimL31 @ 0eb8d6d
d9beuD/oQ6-mtp oQ6MLX150.0download—————no quality claimL27 @ fe8b2e3
d9beuD/oQ6e-mtp oQ6eMLX150.0download—————no quality claimL31 @ fb34e33
d9beuD/oQ8-mtp oQ8MLX195.0download—————no quality claimL26 @ 24a21ad
d9beuD/oQ8e-mtp oQ8eMLX195.0download—————no quality claimL30 @ 3db9c56
dabsLabs/NVFP4 NVFP4 experts-onlyNVFP4 (compressed-tensors nvfp4-pack-quantized, W4A16)174.0download————expert_rel_err_vs_bf16 0.0869; expert_max_rel_err_vs_bf16 0.0870; rel_err_vs_qwen_fp8 0.0266 (48 samples); sq_err_change_vs_rtn −16.11 %vs BF16L71 @ dc1c235
ddalcu/MLX-Serve-8bit 8bitMLX (mlx-serve pack)195.0download—————no quality claimL32 @ f61a712
ddalcu/MLX-Serve-iQ-MLX-3.3bpw iQ-MLX 3.3bpwMLX52.00resident—85.6%——prob_mass 0.732vs BF16L56 @ 4fabeb5
ddalcu/MLX-Serve-iQ-MLX-4.7bpw iQ-MLX-4.7bpwMLX (mlx-serve)75.30resident0.126———top1:all 88.6; top1:agent 89.0; top1:code 96.0; top1:math 90.0; top1:prose 81.2; kld_median 0.0052vs BF16L62 @ a0e89db
ddalcu/MLX-Serve-mixed-4-8bit mixed-4-8bitMLX75.00resident—87.9%——kld_median 0.0066reference not statedL23 @ 788c07a
dealignai/ABLITERATED-NVFP4 modified weightsABLITERATED NVFP4NVFP4 (ModelOpt FP4)—————MMLU (2,280 Q) 81.93 %; harmbench_real_harm_compliance 100 %other or no referenceL71 @ be794b9
dealignai/CYBERSECURITY-NVFP4 modified weightsCYBERSECURITY NVFP4NVFP4—————MMLU 81.27%; MMLU delta -0.84 pp; hard multi-step accuracy 98.3%; reasoning_trace_median 422 tok (0.97×)other or no referenceL134 @ 15ef113
Delgon/PLE-GGUF PLE-BF16 (reference)GGUF (PLE sidecar tables only)102.4download—————other or no referenceL37 @ 5de86c8
Delgon/PLE-GGUF PLE-Q8_0GGUF (PLE sidecar tables only)54.40download————ppl_ratio:held-out 1.0019 +- 0.0013; mean_kld:held-out 0.0162 +- 0.0004; kld99:held-out 0.217; top1_pct:held-out 96.20 +- 0.10 %; ppl_ratio:wikitext-2 0.9987 +- 0.0010; mean_kld:wikitext-2 0.0122 +- 0.0002; kld99:wikitext-2 0.150; top1_pct:wikitext-2 96.44 +- 0.10 %; row_rel_rms_vs_bf16 0.0053other or no referenceL72 @ 5de86c8
Delgon/PLE-GGUF PLE-IQ4_NLGGUF (PLE sidecar tables only)28.80download————ppl_ratio:held-out 1.0022 +- 0.0016; mean_kld:held-out 0.0255 +- 0.0005; kld99:held-out 0.352; top1_pct:held-out 95.02 +- 0.11 %; ppl_ratio:wikitext-2 1.0044 +- 0.0014; mean_kld:wikitext-2 0.0239 +- 0.0004; kld99:wikitext-2 0.289; top1_pct:wikitext-2 94.71 +- 0.11 %; row_rel_rms_vs_bf16 0.076other or no referenceL73 @ 5de86c8
densohax/heretic2-W4A16-FP8PLE modified weightsheretic-2 W4A16 + FP8 PLE + MTP3AutoRound W4A16 (GPTQ packing) + FP8 PLE124.8download—————no quality claimL101 @ 705a2ce
devan-carlin/W4A16 W4A16compressed-tensors W4A1677.00trunk only—————no quality claimL30 @ 40b8f18
DevelopingDad/Uncensored-NVFP4-DGX-Spark modified weightsUncensored NVFP4 (OrcaRouter)NVFP4183.5download—————no quality claimL43 @ 6773ecf
DevQuasar/Qwen.GGUF GGUF (variants not listed)GGUF——————no quality claim@ ef7dd83
DhruvalLabs/GGUF Q4_K_MGGUF111.0download—————no quality claimL33 @ c257284
DJLougen/Mooney modified weightsMooney (PQ2_0)GGUF (PQ2_0 ternary experts)92.00download————suite_score 88.71%; suite_run1 475/536 (88.62%); suite_run2 476/536 (88.81%); retained_pct 95.5%; needle 30/30; top1_pct:long_text ≈72%; kld_top20:long_text ≈0.27vs BF16L140 @ 84f867d
donught/NVFP4-dense NVFP4-denseNVFP4171.0download————GSM8K (n=200) 98.0 %; mean_completion_tokens_gsm8k 450.9other or no referenceL38 @ 8ce6fc7
doth4580/EXL3-3.05bpw EXL3 3.05bpw_h5_ng5EXL385.00download—————no quality claimL32 @ 2f41a00
doth4580/EXL3-4.05bpw 4.05bpw_h6_ng6EXL3108.0download—————no quality claimL34 @ e2e53fd
drluoto/AgenticRequant-Q5K-GGUF AgenticRequant (trunk Q5_K)GGUF92.30download——4.7823 ± 0.0295+0.91 %—other or no referenceL102 @ d69cf60
drluoto/MTP-GGUF bf16GGUF (MTP draft head only)7.8download—————no quality claimL47 @ ac875a9
drluoto/MTP-GGUF Q8_0GGUF (MTP draft head only)4.1download—————no quality claimL44 @ ac875a9
drluoto/MTP-GGUF Q8_0-frspec-65kGGUF (MTP draft head only)3.6download—————no quality claimL43 @ ac875a9
drluoto/MTP-GGUF Q4_K_MGGUF (MTP draft head only)2.8download—————no quality claimL45 @ ac875a9
drluoto/MTP-GGUF Q5_K-frspec-65kGGUF (MTP draft head only)2.7download—————no quality claimL42 @ ac875a9
drluoto/MTP-GGUF Q5_K-frspec-65k-bf16pathGGUF (MTP draft head only)2.7download—————no quality claimL46 @ ac875a9
drowzeys/keys-ablit-Mia-Single-Spark-only modified weights(no variants listed)not stated——————no quality claim@ b3b60a4
Dyluhn/R9V-IQ4_XS UD-IQ4_XS (unsloth, rebundled)GGUF (bundle)——————no quality claim@ f48e978
Dyluhn/exl3 EXL3 experts containerradiance .rad (EXL3 experts)120.5download0.0467 mean_kl94.0%4.0507 (×0.9934)—kl_delta_vs_baseline −12.9%; 95% CI of the difference [−0.0102, −0.0041]vs BF16L42 @ 473882c
dzannotti/MTP-GGUF MTP-BF16GGUF (MTP draft head)7.8download—————no quality claimL36 @ 0b2551d
dzannotti/MTP-GGUF MTP-Q4_K_MGGUF (MTP draft head)2.5download—————no quality claimL35 @ 0b2551d
EasiiX/MTP-Strix-Halo-GGUF mtp-Qwen3.8-Flash-Next-hc-Q8_0.ggufGGUF (MTP sidecar only)4.1download—————no quality claimL32 @ 896756c
EasiiX/MTP-Strix-Halo-GGUF mtp-Qwen3.8-Flash-Next-Q8_0.gguf (deprecated)GGUF (MTP sidecar only)4.1download—————no quality claimL33 @ 896756c
edougawa/NVFP4-PLE-MMap-MTP-Spark NVFP4 body + mmap NVFP4 PLENVFP4 (ModelOpt W4A16_NVFP4)75.05trunk only————ple_row_cosine cosine 0.995540; body_cosine_mean cosine mean 0.995620; body_cosine_min min 0.995146vs BF16L88 @ c3e6123
edougawa/W4A16-PLE4-MTP-Spark W4A16-PLE4compressed-tensors W4A1694.91download—————no quality claimL41 @ 9f10435
eemin/halogen-copy w4b checkpointhalogen .hgn124.1download—————no quality claimL47 @ c01ee53
eemin/halogen-copy w4b overlay sidecarhalogen .hgn2.6download—————no quality claimL48 @ c01ee53
EigenLabs/MLX-4bit-mtp MLX 4-bit mtpMLX106.3download—————no quality claimL36 @ e555958
EliovpAI/W3A8-Paiton-RDNA4 INT3 (W3A8)Paiton (W3A8)116.4download—90.74 %——kl_free 0.0604; kl_routing_locked 0.0413; served_kl_gate 0.0616; nll_doc_level 1.7671 (+0.0097)vs BF16L117 @ e5cee0b
EliovpAI/W4A8-Paiton-RDNA4 4-bit as served (--mode 4bit)Paiton W4A8 (INT4/INT3 hybrid experts)133.0download—93.36 %——GSM8K 96.8 %; HumanEval pass@1 96.3 %; MMLU-Pro 83.7 %; ppl:wikitext-2 512 4.629; ppl:wikitext-2 4096 3.961; ppl:unseen 512 6.954; ppl:unseen 4096 6.187; kl_free/locked 0.0306 / 0.0166; nll:long docs 1.7603vs BF16L44 @ 5e4a8db
evsinlb/oQ4e-fp16-mtp oQ4e-fp16-mtpMLX (oMLX oQ4e)107.2download————agentic_battery 3/3other or no referenceL37 @ d2ec554
fabiogreter/lily-q4 lily-q4lily qwen4_exp-affine-v1 (safetensors, engine-specific)105.2download—————no quality claimL43 @ 28ed324
fabiogreter/lily-q4-xl lily-q4-xllily qwen4_exp-affine-v1 (safetensors, engine-specific)106.9download————tool_call_continuation:opencode 56K turn 23 of 30other or no referenceL44 @ 0f94842
FalconZypp/AWQ-INT4 AWQ-INT4AWQ188.3download—————no quality claimL35 @ 2b1c486
FenomAI/GSQ-RCO-GGUF IQ3_SGGUF83.60download————AIME25 100.00; GPQA-Diamond 92.93; LiveCodeBench v6 86.86; task_avg 93.26vs BF16L127 @ 904afcb
FenomAI/GSQ-RCO-GGUF IQ3_XXSGGUF75.80download————zs_avg 77.23; recovery_pct 100.4%; AIME25 100.00; GPQA-Diamond 91.41; LiveCodeBench v6 86.29; task_avg 92.57vs BF16L126 @ 904afcb
FenomAI/GSQ-RCO-GGUF IQ2_XSGGUF68.00download————zs_avg 77.16; recovery_pct 100.3%; AIME25 96.67; GPQA-Diamond 87.37; LiveCodeBench v6 83.43; task_avg 89.16vs BF16L125 @ 904afcb
FenomAI/GSQ-RCO-GGUF Q2_0GGUF66.40download————zs_avg 78.00; recovery_pct 101.4%; AIME25 96.67; GPQA-Diamond 89.39; LiveCodeBench v6 81.14; task_avg 89.07vs BF16L124 @ 904afcb
FINAL-Bench/POCKET-GGUF Q4_K_MGGUF119.2download——6.030 ± 0.103——other or no referenceL94 @ cb1c47a
FINAL-Bench/POCKET-GGUF Q4_K_M-v2GGUF99.86download——6.080 ± 0.104——other or no referenceL95 @ cb1c47a
fuutott/MLX-t5-imatrix-q3down-ple8 T5 + imatrix gate/up, Q3 down (this checkpoint)MLX (oMLX fork, ternary)97.06download0.48976.1 %4.35 (× 1.32)—median_kld 0.187; MMLU (5-shot) 249 (83.0 %); GSM8K 90 (90.0 %); TruthfulQA MC1 180 (90.0 %); CMMLU (0-shot) 68 (68.0 %); bench_total 587 (83.9 %); HumanEval 152 (92.7 %); MBPP 390 (78.0 %)other or no referenceL107 @ 15ec9e8
gabmiranda/oQ2.5e-mtp oQ2.5eMLX——————no quality claim@ e372f69
gabmiranda/oQ2.7e-mtp oQ2.7eMLX——————no quality claim@ 79f0ee1
gabmiranda/oQ2e-mtp oQ2eMLX——————no quality claim@ 9a235a2
gabmiranda/oQ3.5e-mtp oQ3.5eMLX——————no quality claim@ cf29543
gabmiranda/oQ3e-mtp oQ3e-mtpMLX——————no quality claim@ d1453d3
gabmiranda/oQ4e-mtp oQ4eMLX——————no quality claim@ e3ce05a
gabmiranda/oQ5e-mtp oQ5eMLX——————no quality claim@ cf2679f
gabmiranda/oQ6e-mtp oQ6eMLX——————no quality claim@ 0d4741f
gabmiranda/oQ8e-mtp oQ8eMLX——————no quality claim@ 7453cc0
garnermccloud/MLX-SSD-Stream SSD Stream (Q4 experts, BF16 core)MLX (mlx-serve pack)127.0download—————no quality claimL58 @ 9a624d9
GBP-DE/oQ4e-WLSRefit-MTP oQ4e WLSRefitMLX (oMLX oQ)——————no quality claim@ 3f3eb3a
GBP-DE/oQ5e-WLSRefit-MTP oQ5e WLSRefit MTPMLX (oMLX oQ5e)——————no quality claim@ 559775f
GBP-DE/oQ5e-mtp oQ5eMLX125.7download————test:8-person constraint puzzle (thinking off) Exact solution; test:five-trigger state machine (thinking on) Two synthetic post-hoc verification values (checksum and signature) were calculated incorrectly; test:Python debugging (thinking on) coding failure despite a correct analysis and correct expected results; test:four-tank state machine Exact; test:Boolean/SAT Exactother or no referenceL136 @ 387311f
GCSA-AiLab/FP8-Abliterated-GGUF modified weightsQ4_KMGGUF——————no quality claim@ 881b6c5
GCSA-AiLab/FP8-Abliterated-GGUF modified weightsQ6_KGGUF——————no quality claim@ 881b6c5
GCSA-AiLab/FP8-Abliterated-GGUF modified weightsQ8GGUF——————no quality claim@ 881b6c5
ggml-org/GGUF Q8_0GGUF162.6download (HF listing)—————no quality claim@ 0a61d2b
ggml-org/GGUF IQ4_NLGGUF102.0download (HF listing)—————no quality claim@ 0a61d2b
ggml-org/GGUF Q2_KGGUF83.13download (HF listing)—————no quality claim@ 0a61d2b
glyd/hummingbird preciseGlyd proprietary format82.00download————GSM8K 148 / 150; MMLU-Pro 124 / 150; Star Drift playable 3 of 6other or no referenceL89 @ a101fa6
glyd/hummingbird fastGlyd proprietary format68.00download————GSM8K 148 / 150; MMLU-Pro 118 / 150; Star Drift playable 1 of 6other or no referenceL89 @ a101fa6
gpillon/trellis-a25-ignis trellis a25 (2.5 bpw experts)trellis (EXL3/QTIP family, .ninfer)71.76download0.164 kld:prose———kld:chat 0.052; top1:chat 94.0%; kld:code 0.116; top1:code 89.0%; top1:prose 81.9%; kld:py 0.027; top1:py 98.1%; kld:math 0.047; top1:math 95.8%; kld:en 0.106; top1:en 88.0%; kld:it 0.050; top1:it 91.3%; kld:de 0.151; top1:de 86.1%; kld:ja 0.134; top1:ja 84.2%; kld:zh 0.083; top1:zh 86.2%; kld:mmlu 0.123; top1:mmlu 89.8%; MMLU-Pro proxy (281) 70.1%vs BF16L42 @ f75e492
GreenBitAI/4bit-paged 4bit-pagedMLX (gbx_lm paged containers)113.2download————logits_vs_source exact on 5 prompt(s) to 160 tokens; 1 longer differ by at most 6.8other or no referenceL132 @ e98b855
groxaxo/Q2_0-H64-S4-GGUF Q2_0 H64-S4GGUF (Hadamard-rotated, patched runtime)68.90download————smoke_prompts_thinking_off 21 of 27; retry_thinking_on all six passed; recon_error_vs_direct_Q2 2.91% lowerother or no referenceL38 @ 5eea3d5
groxaxo/exl3-3.50bpw_hq_h6_ng6 3.50bpw_hq_h6_ng6EXL393.00download—————no quality claimL40 @ 5ebeaa0
groxaxo/exl3-unsloth-shaped-h8-mtp5 EXL3 3.82 h8/mtp5/ng6EXL3104.0download————Math (exact numeric) 9/10; Logic / reasoning 4/5; Code (executed) 5/5; Misc (JSON, Spanish ×2) 5/5; Needle in haystack (~20k, ~60k) 2/2; Overall 23/25other or no referenceL180 @ 4972575
Guile/GGUF bf16GGUF354.0download—————no quality claimL44 @ 75862fa
Guile/GGUF Q8_0GGUF188.3download—————no quality claimL45 @ 75862fa
Guile/GGUF Q6_KGGUF168.1download—————no quality claimL46 @ 75862fa
Guile/GGUF Q5_K_LGGUF151.1download—————no quality claimL47 @ 75862fa
Guile/GGUF Q4_K_LGGUF139.3download—————no quality claimL48 @ 75862fa
Guile/GGUF Q5_K_MGGUF134.7download—————no quality claimL49 @ 75862fa
Guile/GGUF Q5_K_SGGUF128.1download—————no quality claimL50 @ 75862fa
Guile/GGUF Q4_K_MGGUF119.6download—————no quality claimL51 @ 75862fa
Guile/GGUF Q3_K_XLGGUF119.3download—————no quality claimL52 @ 75862fa
Guile/GGUF Q4_K_SGGUF113.1download—————no quality claimL53 @ 75862fa
Guile/GGUF Q4_1GGUF111.3download—————no quality claimL54 @ 75862fa
Guile/GGUF Q2_K_LGGUF107.2download—————no quality claimL55 @ 75862fa
Guile/GGUF Q4_0GGUF100.6download—————no quality claimL56 @ 75862fa
Guile/GGUF IQ4_NLGGUF100.3download—————no quality claimL57 @ 75862fa
Guile/GGUF IQ4_XSGGUF97.68download—————no quality claimL58 @ 75862fa
Guile/GGUF IQ3_MGGUF93.23download—————no quality claimL59 @ 75862fa
Guile/GGUF Q3_K_LGGUF93.19download—————no quality claimL60 @ 75862fa
Guile/GGUF Q3_K_MGGUF91.96download—————no quality claimL61 @ 75862fa
Guile/GGUF IQ3_XSGGUF91.95download—————no quality claimL62 @ 75862fa
Guile/GGUF Q3_K_SGGUF89.44download—————no quality claimL63 @ 75862fa
Guile/GGUF IQ3_XXSGGUF88.02download—————no quality claimL64 @ 75862fa
Guile/GGUF Q2_KGGUF80.93download—————no quality claimL65 @ 75862fa
Guile/GGUF IQ2_MGGUF80.37download—————no quality claimL66 @ 75862fa
Guile/GGUF IQ2_SGGUF77.83download—————no quality claimL67 @ 75862fa
Guile/GGUF IQ2_XSGGUF77.74download—————no quality claimL68 @ 75862fa
Guile/GGUF IQ2_XXSGGUF75.19download—————no quality claimL69 @ 75862fa
Guile/GGUF IQ1_MGGUF72.01download—————no quality claimL70 @ 75862fa
Guile/GGUF IQ1_SGGUF70.10download—————no quality claimL71 @ 75862fa
Hagwell/GGUF not statedGGUF——————no quality claim@ 339aed9
haihengh/125B-finch-4bit finch 4-bitFinchMoE .finch (custom streaming layout)162.0download—————no quality claimL57 @ 2464910
halt95/W4A16-Merlin W4A16-MerlinW4A16 (INT4 AutoRound experts) + INT8 GDN + FP8 PLE123.5download————GSM8K 198/200; structured output 80/80vs BF16L45 @ cec53e9
HamboneLabs-AI/uint3-g64 uint3-g64uint3 g64 (ModelOpt mixed)55.30trunk only————GSM8K 0.9650; MATH-500 0.9000; HumanEval 0.8780; bench_pooled 0.9167; delta_pooled −2.66 pp; top1_disagreement 0.0639; mean_kl_nats 0.149other or no referenceL144 @ 2da0f88
him0413/Uncensored-Q4_K_M-MTP modified weightsQ4_K_M + MTP (uncensored)GGUF122.1download—————no quality claimL43 @ 54fc513
hstolte/HALOLOOM-INT4-INT8-AWQ Bf16Core-int4-int8Quark mixed INT4/INT8 (AWQ)188.8download—————no quality claimL60 @ dadedd9
hstolte/HALOLOOM-INT4-INT8-AWQ int8Core-int4-int8Quark mixed INT4/INT8 (AWQ)184.6download—————no quality claimL60 @ dadedd9
htdohk/Huihui-abliterated-MTP-GGUF modified weightsUD-Q4_K_XL (unsloth series)GGUF——————no quality claim@ 83eeb1c
htdohk/Huihui-abliterated-MTP-GGUF modified weightsSwift seriesGGUF——————no quality claim@ 83eeb1c
huihui-ai/Huihui-abliterated-GGUF modified weightsunsloth series (e.g. UD-Q4_K_XL)GGUF——————no quality claim@ 06a318d
huihui-ai/Huihui-abliterated-GGUF modified weightsSwift series (from ukisai GSQ-RCO)GGUF——————no quality claim@ 06a318d
huihui-ai/Huihui-abliterated-v2-GGUF modified weightsUD-Q4_K_XLGGUF——————no quality claim@ 43b4c1a
igorls/mixed-NInfer mixed NInfer v2NInfer container (NVFP4/FP8 + INT4 PLE)113.3download—————no quality claimL57 @ ef61b3c
igorls/mixed-NInfer mixed NInfer v3NInfer container (NVFP4/FP8 + INT4 PLE)109.7download—————no quality claimL221 @ ef61b3c
IHaveNoClueAndIMustPost/Simple-Q3_K Simple-Q3_KGGUF65.39trunk only——4.6587 ± 0.028+1.49%—vs BF16L20 @ 7ee3c66
ilintar/gguf-strix-halo IQ4_NL-PROJFIXGGUF100.0download—————reference unclearL23 @ ba5b0d6
immiq/Orinfer-E8P-Q2A8 E8P Q2A8Orinfer safetensors (E8P)51.30download—————no quality claimL37 @ 4c1ffc2
impulsai/CHImp-Alpha-v1.0 CHImp Alpha v1.0GGUF99.00download0.04193.0 %3.838+0.58 %median_kld 0.011; p99_kld 0.48; HumanEval / HumanEval+ 95.7 % / 94.5 %; MBPP / MBPP+ 91.8 % / 78.6 %; Terminal-Bench Core 52.5 %; BFCL v4 single-turn 73.7 %; BFCL v4 multi-turn 60.5 %other or no referenceL138 @ 17645b9
INCModel3/MXFP4-Mixed-CT-RTN-AutoRound MXFP4-MixedMXFP4—————gsm8k 0.9629; hellaswag 0.6830; mmlu 0.8646; piqa 0.8199; piqa acc_norm 0.8341; hellaswag acc_norm 0.8704; bench_avg acc 0.8326; bench_avg acc_norm 0.8830; niah_multiquery 131k 0.8300; ruler_qa_squad 131k 0.8083vs BF16L28 @ ad4d80a
INCModel3/MXFP4-Mixed-FP8KV-CT-RTN-AutoRound MXFP4-MixedMXFP4-Mixed (llm_compressor)—————gsm8k 0.9280; hellaswag 0.6781; mmlu 0.8646; piqa 0.8270other or no referenceL28 @ fea777a
Inferact/NVFP4 NVFP4NVFP4——————no quality claim@ 103a760
inferencerlabs/MLX-Q4i Q4-INFMLX——————other or no referenceL17 @ 85303bc
inferencerlabs/MLX-Q9 Q9MLX———1.20312—token_accuracy_pct 97.80%; missed_divergence_pct 9.60%other or no referenceL37 @ 627f6d6
infosave/cmf q4tpCMF (cortiq)97.12download————ppl:Rust code 4.111other or no referenceL62 @ d043fa6
infosave/cmf q2tpCMF (cortiq)76.95download————ppl:Rust code 4.763other or no referenceL61 @ d043fa6
Intel/W4A16-AutoRound INT4 (W4A16)AutoRound—————GSM8K 0.9682; MMLU 0.8560; PIQA 0.8215; HellaSwag 0.6873; bench_avg 0.8332; recovery_pct 99.64%vs BF16L17 @ 4c67bf6
intel-ai/MXFP4-BF16Attn-CT-AutoRound MXFP4 experts, BF16 attentioncompressed-tensors (MXFP4)182.6download————GSM8K strict 0.9644; GSM8K flexible 0.9636; MMLU 0.8641; PIQA acc 0.8248; PIQA acc_norm 0.8308; HellaSwag acc 0.6856; HellaSwag acc_norm 0.8726vs BF16L89 @ 2855a00
intel-ai/MXFP4-BF16Attn-MXFP8MTP-CT-AutoRound MXFP4 experts + MXFP8 MTP experts, BF16 attncompressed-tensors (MXFP4/MXFP8)180.1download————GSM8K strict 0.9651; GSM8K flexible 0.9644; MMLU 0.8655; PIQA acc 0.8243; PIQA acc_norm 0.8330; HellaSwag acc 0.6845; HellaSwag acc_norm 0.8708vs BF16L86 @ 7810203
intel-ai/MXFP4-BF16Attn-MXFP8MTP-PLEMXFP8-CT-AutoRound MXFP4 experts + MXFP8 MTP experts + MXFP8 PLE, BF16 attncompressed-tensors (MXFP4/MXFP8)131.5download————GSM8K strict 0.9636; GSM8K flexible 0.9636; MMLU 0.8656; PIQA acc 0.8183; PIQA acc_norm 0.8341; HellaSwag acc 0.6832; HellaSwag acc_norm 0.8728; ple_mxfp8_snr_db 53.6 dBvs BF16L101 @ 9008b69
intel-ai/MXFP4-BF16Attn-PLEMXFP8-CT-AutoRound MXFP4 + PLE MXFP8 (BF16Attn)compressed-tensors (MXFP4 experts + MXFP8 PLE)133.0download————GSM8K (strict / flexible) 0.9636 / 0.9636; MMLU 0.8650; PIQA (acc / acc_norm) 0.8150 / 0.8319; HellaSwag (acc / acc_norm) 0.6834 / 0.8726; ple_snr_db 53.6 dBvs BF16L103 @ b17da51
intel-ai/MXFP4-BF16SharedExperts-CT-AutoRound MXFP4 + BF16 shared expertscompressed-tensors mixed-precision (MXFP4/MXFP8)180.0download————GSM8K strict 0.9666; GSM8K flexible 0.9659; MMLU 0.8643; PIQA acc 0.8145; PIQA acc_norm 0.8275; HellaSwag acc 0.6832; HellaSwag acc_norm 0.8694vs BF16L92 @ a0ed54e
intel-ai/MXFP4-Mixed-CT-AutoRound MXFP4-Mixedcompressed-tensors (MXFP4/MXFP8)180.4download————GSM8K (strict) 0.9621; MMLU 0.8626; PIQA 0.8226; HellaSwag 0.6832; bench_avg 0.8326; delta_avg −0.35 ppvs BF16L73 @ 2c55b10
intel-ai/MXFP8-CT-AutoRound MXFP8MXFP8 (compressed-tensors)240.4download————GSM8K strict 0.9704; GSM8K strict stderr ±0.0047; MMLU 0.8514; MMLU stderr ±0.0029; PIQA 0.8210; HellaSwag 0.6913; bench_avg 0.8335; delta_avg −0.26 pp; delta:MMLU −1.37 ppvs BF16L164 @ 8c4d705
ISTA-DASLab/GSQ-RCO-Coder-GGUF modified weightsGSQ-RCO CoderGGUF58.40download————SWE-bench Verified 75.60; LiveCodeBench v6 86.28; retained:SWE-bench Verified 91.3%; retained:LiveCodeBench v6 98.7%vs BF16L57 @ 5348543
ISTA-DASLab/GSQ-RCO-GGUF IQ3_SGGUF83.60download————AIME25 100.00; GPQA-Diamond 92.93; LiveCodeBench v6 86.86; task_avg 93.26vs BF16L127 @ ed59f92
ISTA-DASLab/GSQ-RCO-GGUF IQ3_XXSGGUF75.80download————zs_avg 77.23; recovery_pct 100.4%; AIME25 100.00; GPQA-Diamond 91.41; LiveCodeBench v6 86.29; task_avg 92.57vs BF16L126 @ ed59f92
ISTA-DASLab/GSQ-RCO-GGUF IQ2_XSGGUF68.00download————zs_avg 77.16; recovery_pct 100.3%; AIME25 96.67; GPQA-Diamond 87.37; LiveCodeBench v6 83.43; task_avg 89.16vs BF16L125 @ ed59f92
ISTA-DASLab/GSQ-RCO-GGUF Q2_0GGUF66.40download————zs_avg 78.00; recovery_pct 101.4%; AIME25 96.67; GPQA-Diamond 89.39; LiveCodeBench v6 81.14; task_avg 89.07vs BF16L124 @ ed59f92
ISTA-DASLab/P48NVFP4-MoESQ modified weightsP48NVFP4-MoESQNVFP4 (W4A4) + paired-4:8 sparsity160.0download————ARC-C 74.74; GSM8K 94.34; HellaSwag 83.54; MMLU 84.56; TQA-MC2 61.84; Winogrande 75.16; bench_avg 79.03 ± 0.15; recovery_pct 96.3 %; LCB-v6 65k 49.03 ± 1.34; LCB-v6 131k 53.54 ± 1.45; SWE-bench Verified 67.0; recovery_pct:LCB131k/SWE 93.9 % / 84.0 %vs BF16L113 @ 80266c6
ivanfioravanti/DS4-IQ2 IQ2XXSImatrix-Q2KDownPad768-MTPGGUF44.81disk_main_only—90.32% top1_pct_vs_bf16——target_nll 0.303570; first_token_match 65/99; hard_smoke_12q 5 correct, 0 wrong, 7 incompletevs BF16L18 @ b8b2039
ivanfioravanti/DS4-Q4 Q4KImatrixExperts-MXFP4Down-...-MTPGGUF74.90disk_main_only—96.6%——target_logprob_mae 0.0431; first_token_match 74/85; avg_target_nll 0.213vs BF16L57 @ 59a55fb
Jakevin/ternary-mlx ternary mixMLX (mlx-ternary mixed)—————kl_nats_per_token 0.167; mcq_correct_of_216 149other or no referenceL30 @ e5f8a6b
jamesrogers/MTP-MXFP4-GGUF MXFP4-ngramBF16-MTP (legacy)GGUF174.8download—————other or no referenceL65 @ 0a9ac88
jamesrogers/MTP-MXFP4-GGUF MXFP4-ngramBF16GGUF172.0download—————other or no referenceL64 @ 0a9ac88
jamesrogers/MTP-MXFP4-GGUF MXFP4-ngramQ8-NextNGGUF126.8download————ppl_delta_pct_vs_bf16_table +0.12%other or no referenceL63 @ 0a9ac88
jamesrogers/MTP-MXFP4-GGUF MXFP4-ngramQ8-MTP (legacy)GGUF126.8download—————other or no referenceL65 @ 0a9ac88
JANGQ-AI/JANGH4 JANGH4MLX (JANGH)134.9download————kl_median:general 0.0017; kl_mean:general 0.079; top1:general 94.36%; kl_median:reasoning 0.0002; kl_mean:reasoning 0.041; top1:reasoning 95.93%; kl_median:tool 0.0042; kl_mean:tool 0.222; top1:tool 88.66%vs BF16L72 @ 28cb56c
jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4 CIRU v5.0GGUF (custom CIRU IU4 format)72.11download0.050030 kl94.482%——tail_ppl 2.04505; hard_tools_per100 80; hermes_pass1_per100 93; hermes_pass2_per100 96vs BF16L187 @ f3f020e
jcbtc/Qwen3.8-Flash-CIRU-STRIX-Orca modified weightsOrca v5.0 (compact)GGUF (CIRU compact expert format)72.11download0.04796695.21%2.04668—nll 0.716220other or no referenceL107 @ 139c9f5
jedisct1/oQ4e-100K-MTP oQ4e (MTP disabled)MLX108.8download————tool_calls:positive all 50 positive cases passed; tool_calls:adversarial_guards all 10 adversarial guardsother or no referenceL49 @ a7b39c7
jedisct1/oQ4e-100K-MTP oQ4e (MTP enabled)MLX—————tool_calls:positive all 50 positive cases still passed; tool_calls:adversarial_guards 9 of 10 guards; mtp_accepted 10,631 of 12,152other or no referenceL148 @ a7b39c7
ji-farthing/ik-llama-GGUF v3 IQ4_KTGGUF88.85download0.057803 kld:english91.183% top1:english3.3893 ppl:english—kld:code 0.055726; top1:code 95.142%; ppl:code 1.7303vs BF16L110 @ 8fb1860
ji-farthing/ik-llama-GGUF v3 IQ3_KTGGUF77.87download0.085269 kld:english89.384% top1:english3.4484 ppl:english—kld:code 0.077747; top1:code 94.653%; ppl:code 1.7522vs BF16L111 @ 8fb1860
ji-farthing/ik-llama-GGUF_v2 IQ4_KT_v2GGUF86.67download0.132805 kld:english86.276% top1:english——kld:code 0.124558; top1:code 93.216%vs BF16L73 @ ca863d1
jlkivey/MTP-PR27836-GGUF mtp-Qwen3.8-Flash-Next-Q8_0 (draft head only)GGUF4.1download—————no quality claimL32 @ 159500b
jmgao/Swift1.5-Flash-Next-mixed-NInfer modified weightsSwift1.5 mixed NInferNInfer mixed (NVFP4/FP8/INT4)——————no quality claim@ 2f012c4
Joakimpalm-Zen/MTP-GGUF MTP-Q8_0 (draft head)GGUF (MTP draft head)4.2download—————no quality claimL28 @ c104e65
jockevaupptaget/MTP-GGUF mtp Q8_0GGUF (MTP draft head only)4.1download—————no quality claimL14 @ 69da733
jomcgi-org/NVFP4-oominf NVFP4 oominf (serve defaults)NVFP4135.0download————GSM8K (1,319) 97.73; GPQA-Diamond (first 118) 88.1; GSM8K (300, t0) 97.33; MMLU-Pro (200) 77.00; RULER 8k+32k 99.23; RULER 95k 98.46vs BF16L90 @ 84010e0
Jon-Nielsen/AWQ-INT4-FP8PLE AWQ-INT4 + FP8 PLEcompressed-tensors AWQ INT4 + FP8 PLE137.1download————ple_shard0_snr_db 31.5 dB; ple_max_rel_err 0.195other or no referenceL18 @ c2aa55f
jpezzulli/OrcaRouter-Uncensored-ModelOpt-NVFP4 modified weightsModelOpt NVFP4NVFP4 (ModelOpt 0.46.0)135.3download—————no quality claimL43 @ f24d2b6
JRamirez-UAB/EXL3-2.25bpw-48GB EXL3 2.25bpwEXL340.46trunk only————mean_kld:internal 0.0532; median_kld:internal 0.0215; mean_kld:public 0.0481; median_kld:public 0.0064; ppl:internal 2.1944; ppl:public 1.7135; agent tasks 30/30; tools opencode 58/60; reasoning xhigh 48/48vs BF16L82 @ 9eb541a
Jrouth/MTP-C2T8-refit-GGUF modified weightsmtp-Qwen3.8-Flash-Next-Q4DRAFT-refit0930.ggufGGUF (MTP draft head only)2.7download—————no quality claimL63 @ 7876390
Jrouth/MTP-C2T8-refit-GGUF modified weightsmtp-Qwen3.8-Flash-Next-Q4DRAFT-refit0930-upstream.ggufGGUF (MTP draft head only)2.7download—————no quality claimL64 @ 7876390
junafinity/Qwen-3.8-Flash-Next-Uncensored-MLX-MXFP4 modified weightsMXFP4 mixedMLX100.5download—————no quality claimL44 @ 6fd6569
Jundot/oQ2.5e-fp16-mtp oQ2.5e-fp16MLX83.80download—————no quality claimL53 @ 9eed92c
Jundot/oQ2.5e-mtp oQ2.5eMLX82.90download—————no quality claimL53 @ ebfb104
Jundot/oQ2.7e-mtp oQ2.7eMLX85.20download—————no quality claimL50 @ 0e751a5
Jundot/oQ3.5e-fp16-mtp oQ3.5e-fp16MLX97.13download—————no quality claimL55 @ 42ce563
Jundot/oQ3.5e-mtp oQ3.5eMLX96.23download—————no quality claimL55 @ e821605
Jundot/oQ3e-fp16-mtp oQ3e-fp16MLX92.10download—————no quality claimL55 @ 14f743e
Jundot/oQ3e-mtp oQ3eMLX91.20download—————no quality claimL55 @ b73babc
Jundot/oQ4e-fp16-mtp oQ4e-fp16MLX107.5download—————no quality claimL54 @ 3ed70e9
Jundot/oQ4e-mtp oQ4e (updated)MLX106.3download————kl:code 0.0776; kl:novel 0.1713vs BF16L34 @ c4ddfb5
Jundot/oQ6e-mtp oQ6eMLX150.5download—————no quality claimL48 @ 816d1eb
Jundot/oQ8e-mtp oQ8eMLX194.9download—————no quality claimL46 @ e6658a3
KaedeTai/Uncensored-oQ4e-mtp-zhTW modified weightsUncensored oQ4e-mtp zhTWMLX (per repo name)——————no quality claim@ c4d04ba
Kandandan/MLX-4bit-MTP 4bit g64 + MTPMLX77.00download—————no quality claimL33 @ 40e15b0
Kashapps/Uncensored-ROCmFP4-FAST-GGUF modified weightsQ4_0_ROCMFP4_FAST (uncensored)GGUF (ROCmFP4, fork-only)94.42download————refusal:Harmful (24) 22 comply; Harmless (12) 11 ok; quality:Quality (8) 6/8other or no referenceL72 @ 1aa1c09
KENIC-1/GGUF-UD-IQ4_XS UD-IQ4_XS (from repo name)GGUF——————no quality claim@ c639d07
KENIC-1/GGUF-UD-Q4_K_XL UD-Q4_K_XL (from repo name)GGUF——————no quality claim@ 6bdfee4
khuasar/REAP-288-Q4g32-Q8-MTPLX modified weightsREAP-288 Q4g32+Q8MLX82.26download————8-task suite (AR) 7/8; 8-task suite (MTP depth 3) 6/8other or no referenceL135 @ 0585867
kingjones777/MTP-Heads-GGUF mtp BF16GGUF (MTP draft heads)7.8download—————no quality claimL92 @ 72805b6
kingjones777/MTP-Heads-GGUF mtp Q8_0GGUF (MTP draft heads)4.1download—————no quality claimL97 @ 72805b6
kingjones777/MTP-Heads-GGUF mtp Q6_KGGUF (MTP draft heads)3.4download—————no quality claimL96 @ 72805b6
kingjones777/MTP-Heads-GGUF mtp Q4_K_MGGUF (MTP draft heads)2.8download—————no quality claimL95 @ 72805b6
kingjones777/MTP-Heads-GGUF mtp Q4_0_ROCMFP4_STRIXGGUF (MTP draft heads)2.3download—————no quality claimL94 @ 72805b6
kingjones777/MTP-Heads-GGUF mtp Q4_0_ROCMFP4_FASTGGUF (MTP draft heads)2.1download—————no quality claimL93 @ 72805b6
kingjones777/ROCmFP4-FAST-GGUF ROCmFP4 FASTGGUF (ROCmFP4)94.42download—————no quality claimL134 @ 7565b1e
kingjones777/ROCmFP4-STRIX-GGUF ROCmFP4 STRIXGGUF (ROCmFP4, Q4_0_ROCMFP4_STRIX)121.8download—————no quality claimL310 @ 7320914
kingjones777/ROCmFP4-STRIX_LEAN-GGUF Q4_0-ROCmFP4-STRIX_LEANGGUF (ROCmFP4)105.8download—————no quality claimL134 @ bf7fbc6
kingjones777/Uncensored-ROCmFP4-FAST-GGUF modified weightsQ4_0_ROCMFP4_FAST (Uncensored)GGUF (ROCmFP4)94.42download————quality_set 6/8; harmless_ok 11; harmful_comply 22other or no referenceL213 @ ce1bddb
kingjones777/Uncensored-ROCmFP4-FAST-imatrix-GGUF modified weightsROCmFP4-FAST imatrixGGUF (ROCmFP4)94.38download——5.0337 ± 0.031−5.9%ppl:4bit_head 5.1075other or no referenceL58 @ 9abf014
kingjones777/Uncensored-ROCmFP4-STRIX_LEAN-GGUF modified weightsQ4_0_ROCMFP4_STRIX_LEAN (uncensored)GGUF (ROCmFP4, fork-only)105.8download————refusal:Harmful (24) 22 comply; Harmless (12) 11 ok; quality:Quality (8) 6/8other or no referenceL129 @ a4d9d7b
kingjones777/Uncensored-ROCmFP4-STRIX_LEAN-imatrix-GGUF modified weightsQ4_0-ROCmFP4-STRIX_LEAN-imatrixGGUF105.8download——4.9865 ± 0.031—ppl_delta_pct_vs_plain_FAST −6.7%other or no referenceL58 @ 503e614
klee100/AutoRound-3bpw-MTP AutoRound 3bpw + MTPAutoRound (GPTQ-format)50.79GPU weights0.1208 kld————vs BF16L30 @ 600715b
kstoyanov99/qwen38-flash-next-nvfp4 NVFP4NVFP4135.0download————GSM8K 97.27; AIME26 pass@1 98.75vs BF16L138 @ ee25dae
Kymidnight/ROCmFP4-FAST-GGUF ROCmFP4-FASTGGUF89.82download——4.6785 +/- 0.02780——vs BF16L76 @ 6dd6500
lackonendes/PAW-125B-FLASH-NEXT-X3.1 PAW-125B X3.1GGUF (TRUSS-only)51.00trunk only————kl:code 0.044; kl:plain-English 0.087; GPQA-Diamond 92.4 (183/198); LiveCodeBench v6 74.9 (131/175, in progress); IFBench 76.7 (115/150); retention:GPQA-Diamond 100.8%; retention:LiveCodeBench v6 81.5% (partial); retention:IFBench 94.3%vs BF16L123 @ 257599e
lee-chang-93/NVFP4-REAP-k448 modified weightsNVFP4-REAP-k448NVFP4 (ModelOpt)124.1download————reap_mass_retained_pct 92.65% retainedother or no referenceL28 @ 8d565c9
leoncca/AWQ-g32 AWQ g32 (2026-10-17)AWQ137.1download—97.06%——kld_top20_tail 0.0207; kld:instructions 0.0278; top1:instructions 95.56%; kld:dialogue 0.0236; top1:dialogue 96.17%; kld:documents_code 0.0194; top1:documents_code 97.49%; kld_bounds [0.020465, 0.317644]vs BF16L136 @ ff2fb53
lesj0610/NVFP4 NVFP4NVFP4186.4download————weight_cos 0.9954; weight_rel_frob_err 0.094vs BF16L170 @ e8045c6
lhoestq/NVFP4-Michel NVFP4NVFP4—————GPQA Diamond 91.5; HLE 35.4; τ²-Bench Telecom 90.1; MMMU Pro 78.3; SciCode 18.8; AA-LCR 74.1; IFBench 81.0; Omniscience 27.6; Terminal-Bench 2.1 82.9other or no referenceL93 @ 81d3bdd
Litwein/REAP320-oQ3e-DWQ-MTP-Vision-MLX modified weightsREAP320 oQ3e DWQ v9MLX70.80download————ce_gap:Reasoning 0.0596; kl20:Reasoning 0.1077; ce_gap:Code 0.0437; kl20:Code 0.0595; ce_gap:Russian chat 0.1561; kl20:Russian chat 0.1136; ce_gap:Knowledge QA 0.0950; kl20:Knowledge QA 0.1483; ce_gap:Agent 0.2235; kl20:Agent 0.1854; ce_gap:Overall 0.0901; kl20:Overall 0.1035; ce_gap:independent_gate 0.0968; kl20:independent_gate 0.0833; ce_gap:vision 0.2533; kl20:vision 0.2245; ce_gap_change_vs_pruned_init −14.9%; kl20_change_vs_pruned_init −11.2%; pruning_gap_recovered_pct 33.6%vs BF16L139 @ 9a16783
Litwein/REAP320-oQ3e-DWQ-MTP-Vision-MTPLX modified weightsREAP320-oQ3e-DWQ MTPLXMLX (MTPLX layout)39.73trunk only—————no quality claimL62 @ 2f7b82f
Litwein/REAP320-oQ3e-fp16-DWQ-MTP-Vision-MLX inherited family evaluationmodified weightsREAP320 oQ3e fp16 DWQ v9MLX (oMLX)71.70download————ce_gap:Overall 0.0901; kl20:Overall 0.1035; kl20_delta_v8_v9:Overall −1.6%; ce_gap:Reasoning 0.0596; kl20:Reasoning 0.1077; ce_gap:Code 0.0437; kl20:Code 0.0595; ce_gap:Russian chat 0.1561; kl20:Russian chat 0.1136; ce_gap:Knowledge QA 0.0950; kl20:Knowledge QA 0.1483; ce_gap:Agent 0.2235; kl20:Agent 0.1854; ce_gap:independent gate 0.0968; kl20:independent gate 0.0833; ce_gap:real-image vision 0.2533; kl20:real-image vision 0.2245; ce_gap_reduction_vs_pruned_init 14.9%; kl20_reduction_vs_pruned_init 11.2%; pruning_gap_recovered_pct 33.6%; fp16_exact_values_pct 99.99973%; fp16_overflow_values 0; fp16_flushed_to_zero 148vs BF16L194 @ 46a83dc
Litwein/REAP320-oQ3e-fp16-DWQ-MTP-Vision-MTPLX modified weightsREAP320 oQ3e fp16 DWQ MTPLXMLX (MTPLX layout)—————fp16_exact_values_pct 99.99973%; fp16_overflow_values 0; fp16_flushed_to_zero 598vs BF16L92 @ 0bb3ca6
lmstudio-community/GGUF Q8_0GGUF188.2download (HF listing)—————no quality claim@ 158fc82
lmstudio-community/GGUF Q6_KGGUF167.6download (HF listing)—————no quality claim@ 158fc82
lmstudio-community/GGUF Q4_K_MGGUF119.2download (HF listing)—————no quality claim@ 158fc82
local-inference-lab/NVFP4 modified weightsNVFP4/MXFP8 QADNVFP4105.2download————AA-LCR 79.4; GPQA Diamond 91.41; IFBench 82.31; Tool Eval 95other or no referenceL110 @ 6c01ac3
LOKSUN-158/SamQuant-Uncensored-MLX-Serve-Mix-C58 modified weightsC58MLX (mlx-serve mixed)92.00download—85.40%——acceptance:tool_orchestration 3/8vs BF16L87 @ a6692a8
Lottolabs/Coder-P150 modified weightsCoder GPTQ-BFP4 experts (expert files only)BFP4 (Tenstorrent)59.80download0.038 kl——+0.09%top1_agreement 0.937; GSM8K first 50 45/50other or no referenceL34 @ 69e738b
Lottolabs/Coder-P150 modified weightsfull model (IQ3_S) card residentsBFP4 (Tenstorrent)—.0279 kl——-0.66%top1_agreement .942; GSM8K-50 45/50other or no referenceL73 @ 69e738b
lovedheart/NVFP4-FP8 NVFP4NVFP4 (+FP8 PLE)135.0download————GSM8K 97.27; AIME26 pass@1 98.75vs BF16L225 @ a9786bf
lovedheart/NVFP4-FP8-Pruned-RTXPRO-6000 modified weightsNVFP4-FP8 pruned (448 experts)NVFP4+FP8—————GSM8K 97.27; AIME26 pass@1 98.75vs BF16L195 @ 0e63a89
lovedheart/NVFP4-W4A16-ATTN-FP8-MTP-NVFP4 NVFP4-W4A16 4o6 + FP8 attn, MTP v4NVFP4 (ModelOpt mixed: NVFP4 W4A16 experts + FP8 block attention)122.0download—————no quality claimL71 @ f748827
Lucebox/GGUF MTP shared Q8_0GGUF2.8download—————no quality claimL28 @ f048355
Lucebox/GGUF UD-Q4_K_XL (unsloth mirror)GGUF——————no quality claim@ f048355
LunarMoonar/Huihui-abliterated-GGUF modified weightsUD-Q4_K_XL (unsloth series)GGUF——————no quality claim@ e0db1ef
LunarMoonar/Huihui-abliterated-GGUF modified weightsSwift seriesGGUF——————no quality claim@ e0db1ef
maclocal/AFM-MLX-4bit AFM MLX 4-bitMLX (AFM)——————no quality claim@ 2ac1a93
madlabs-ai/NVFP4-MTP-BF16 NVFP4 body + BF16 MTPNVFP4 (ModelOpt)——————no quality claim@ 85bee1e
MagneticLab/NVFP4 NVFP4 (full download)NVFP4 (ModelOpt) + NVFP4 PLE sidecar200.0download————ple_sidecar_worst_shard_rel_err 0.0896reference not statedL61 @ e30ead7
Manav1011/NVFP4 NVFP4NVFP4135.0download————GSM8K 97.27; AIME26 pass@1 98.75vs BF16L120 @ f86fa72
manjunathshiva/tq4a-tq2e-g64 tq4a-tq2e-g64MLX (TurboQuant)55.83download——9.31—battery:thinking_disabled_5_seeds 5/5 clean; battery:thinking_enabled_close_within_5000 7/15other or no referenceL224 @ 34811dc
manuuuel/mica-q4 mica q4GGUF (Mica)177.3download————mean_kld:All 0.0399; kld_p99:All 0.585; top1_pct:All (authors' code) 94.00%; ppl_delta_pct:All (authors' code) +0.14%; mean_kld:English prose 0.0542; mean_kld:Code 0.0080; mean_kld:Chat 0.0567; mean_kld:Other languages 0.0407; nats_per_token:All (Mica) 1.5510; ppl_delta_pct:All (Mica) +0.12%; top1_pct:All (Mica) 93.98%vs BF16L116 @ 354af7a
matteiuspi/W4A16-G128-300i W4 G128 (native W4A8 path)ModelSlim INT4 W4 g128 (Ascend 310P)181.6download————quality_gate_228q_zero_shot 206/228; GPQA Diamond (AISBench Answer: X) 140/198 (70.71%); gpqa_final_response_correct 114/198 (57.58%); gpqa_hit_output_limit 81/198other or no referenceL94 @ d32978c
matteiuspi/W8A8-DYNAMIC-300i W8A8 DynamicAscend ModelSlim W8A8_DYNAMIC (ascend_v1)240.1download——12.2393—mean_nll 2.5046; GPQA Diamond first-5 correct 5 / 5; GPQA Diamond first-5 acc 100%; smoke:semantic_tool_7 6/7vs BF16L321 @ 27e5cdc
Maximilian228/NVFP4-GPTQ-Strata NVFP4 GPTQNVFP474.09RAM in use0.03893.6%——expert_rel_err_vs_bf16 0.00561vs a Q8_0L36 @ 379ff0b
Maximilian228/NVFP4-Q8-Mix-Strata NVFP4 + Q8_0 mixNVFP4+Q8_0 (Strata pack)171.6download0.0215 mean_kl95.0%——top1:thinking 97.4%; kl:thinking 0.0058; ppl_vs_ref 100.0%other or no referenceL41 @ 9ac833e
Maximilian228/Q8_0-Strata Q8_0 (qwen-q8)Strata pack (Q8_0 experts)189.2download—100% (the reference)——expert_rel_mse_vs_bf16 0.00006vs a Q8_0L37 @ d8fc7df
mazinb/Uncensored-NVFP4 modified weightsNVFP4NVFP4——————no quality claim@ f2c21eb
mbehr90/fp8 FP8FP8 (compressed-tensors, FP8_DYNAMIC)237.1download————GSM8K (200 q, chat) 0.965; mean_answer_chars 1146vs BF16L29 @ e70592b
mbehr90/nvfp4 NVFP4NVFP4A16 (compressed-tensors)182.8download————GSM8K (200 q, chat) 0.960; mean_answer_chars 1087vs BF16L30 @ d28eef6
mduca771/NG8-GGUF UD-Q3_K_XL-NG8GGUF115.6download—————no quality claimL51 @ c516e96
mduca771/NG8-GGUF UD-IQ3_XXS-NG8GGUF107.6download—————no quality claimL50 @ c516e96
meldlabs/Meld-Turbo-Quality-GGUF Quality-2.6bpwGGUF (EXL3 expert tensors)74.90download——2.234—GSM8K (200) 96.5%; HumanEval (164) 97.6%; MMLU (200) 87.5%vs a Q8_0L44 @ ecbcece
mensaprodigy/REAP-288-oQ4e-BF16-MTP-PLE modified weightsREAP-288 oQ4eMLX—————HumanEval (100 of 164) 91.0%; LiveCodeBench (100 of 1,054) 56.0%other or no referenceL41 @ 0ee18f7
mensaprodigy/REAP-384-mlx-mtp modified weightsREAP-384 oQ4e + native MTPMLX (oMLX oQ4e)——————no quality claim@ ae3a504
mensaprodigy/REAP-384-oQ4e-BF16-MTP-PLE modified weightsREAP-384 oQ4eMLX (oMLX)—————HumanEval (Thinking No) 91 / 100; HumanEval (Thinking Yes) 95 / 100; LiveCodeBench (Thinking No) 55 / 100; LiveCodeBench (Thinking Yes) 56 / 100other or no referenceL44 @ 57cccc3
mensaprodigy/REAP-384-oQ5e-BF16-MTP-PLE modified weightsREAP-384 oQ5eMLX—————HumanEval (100, Thinking No) 97.0%; HumanEval (100, Thinking Yes) 96.0%; LiveCodeBench (100, Thinking No) 61.0%; LiveCodeBench (100, Thinking Yes) 62.0%other or no referenceL75 @ ff32837
meshapplied/NVFP4-Latest NVFP4NVFP4 (ModelOpt, per HF tags)——————no quality claim@ 13b21e1
MESHIVEAI/NVFP4-W4A16 NVFP4-W4A16NVFP4 (ModelOpt, W4A16)181.5download—————no quality claimL34 @ d58bb7a
Minachist/INT4-Mixed-AutoRound AutoRound mainAutoRound (compressed-tensors)—0.014705———median_kld 0.000872; IFBench, prompt-level loose 81.0; GPQA Diamond 90.4; LiveCodeBench v6, pass@1 92.4vs BF16L57 @ 9bea61d
Minachist/INT4-Mixed-AutoRound AutoRound new-mainAutoRound (compressed-tensors)—0.011680———median_kld 0.000620vs BF16L58 @ 9bea61d
mk2878d/GGUF AD-5.00bpw-Q5_K_M-M64GGUF110.5download0.083789.55%—1.026—vs BF16L84 @ 69652fb
mk2878d/GGUF AD-4.27bpw-Q4_K_M-M64GGUF92.90download0.084289.49%—1.026—vs BF16L83 @ 69652fb
mk2878d/GGUF AD-3.84bpw-IQ4_XS-M64GGUF84.90download0.227782.68%—1.102—vs BF16L82 @ 69652fb
mkinc/rcO4-GGUF rcO4GGUF109.6download—————no quality claimL17 @ b98c43c
mlx-community/4bit 4bit gs32MLX111.5download (HF listing)—————no quality claim@ 07b5dc6
mlx-community/4bit-mtp MTP head shardMLX1.7download—————no quality claimL23 @ 825df62
mlx-community/4bit-mtp 4bit + MTP shardMLX——————no quality claim@ 825df62
mlx-community/OptiQ-2bit OptiQ-2bitMLX (mlx-optiq)81.00download—————no quality claimL45 @ 52a21f4
mlx-community/oQ6e-mtp oQ6eMLX——————no quality claim@ e171af8
mlx-community/oQ8e-mtp oQ8eMLX——————no quality claim@ 2a9de02
MonumentalSystems/control-vectors control vectors (verbosity, comment-density)GGUF control vectors——————no quality claim@ 455a5a1
mradermacher/GGUF Q8_0GGUF188.3download—————no quality claimL54 @ 68d514c
mradermacher/GGUF Q6_KGGUF167.7download—————no quality claimL53 @ 68d514c
mradermacher/GGUF Q5_K_MGGUF134.2download—————no quality claimL52 @ 68d514c
mradermacher/GGUF Q5_K_SGGUF127.8download—————no quality claimL51 @ 68d514c
mradermacher/GGUF Q4_K_MGGUF119.3download—————no quality claimL50 @ 68d514c
mradermacher/GGUF Q4_K_SGGUF111.9download—————no quality claimL49 @ 68d514c
mradermacher/GGUF IQ4_XSGGUF98.50download—————no quality claimL48 @ 68d514c
mradermacher/GGUF Q3_K_LGGUF96.80download—————no quality claimL47 @ 68d514c
mradermacher/GGUF Q3_K_MGGUF94.20download—————no quality claimL46 @ 68d514c
mradermacher/GGUF Q3_K_SGGUF88.90download—————no quality claimL45 @ 68d514c
mradermacher/GGUF Q2_KGGUF80.50download—————no quality claimL44 @ 68d514c
mradermacher/i1-GGUF i1-Q6_KGGUF167.7download—————no quality claimL58 @ 9547b1c
mradermacher/i1-GGUF i1-Q5_K_MGGUF134.2download—————no quality claimL57 @ 9547b1c
mradermacher/i1-GGUF i1-Q5_K_SGGUF127.8download—————no quality claimL56 @ 9547b1c
mradermacher/i1-GGUF i1-Q4_K_MGGUF119.3download—————no quality claimL55 @ 9547b1c
mradermacher/i1-GGUF i1-Q4_K_SGGUF111.9download—————no quality claimL54 @ 9547b1c
mradermacher/i1-GGUF i1-Q4_1GGUF111.1download—————no quality claimL53 @ 9547b1c
mradermacher/i1-GGUF i1-Q4_0GGUF100.4download—————no quality claimL52 @ 9547b1c
mradermacher/i1-GGUF i1-IQ4_XSGGUF97.60download—————no quality claimL51 @ 9547b1c
mradermacher/i1-GGUF i1-Q3_K_LGGUF96.80download—————no quality claimL50 @ 9547b1c
mradermacher/i1-GGUF i1-Q3_K_MGGUF94.20download—————no quality claimL49 @ 9547b1c
mradermacher/i1-GGUF i1-IQ3_MGGUF89.60download—————no quality claimL48 @ 9547b1c
mradermacher/i1-GGUF i1-IQ3_SGGUF89.00download—————no quality claimL47 @ 9547b1c
mradermacher/i1-GGUF i1-Q3_K_SGGUF88.90download—————no quality claimL46 @ 9547b1c
mradermacher/i1-GGUF i1-Q2_KGGUF80.50download—————no quality claimL45 @ 9547b1c
msuiche/abliterated-cyber-GLP-47 modified weightsabliterated-cyber-GLP-47not stated——————no quality claim@ bb05ede
mudler/APEX-GGUF APEX-I-CompactGGUF85.10download—————vs BF16L41 @ d0ab361
mudler/APEX-GGUF APEX-I-MiniGGUF78.70download——4.4354 +/- 0.085+2.9%—vs BF16L91 @ d0ab361
mudler/APEX-GGUF APEX-I-NanoGGUF73.00download——4.6659 +/- 0.093+8.2%—vs BF16L92 @ d0ab361
munekazu/exl3-CMP170HX-64GB EXL3 mix (20 x 4-bit layers)EXL3—————HumanEval (base tests) 157 / 164 (95.7%); HumanEval+ 152 / 164 (92.7%); GSM8K (first 300) 293 / 300 (97.7%)other or no referenceL61 @ b429d27
mxxm-t/MTP-GGUF qwen4exp-mtp-head-BF16GGUF (MTP draft head only)7.8download————draft_acceptance_pct 75.1%vs BF16L38 @ c7c423f
mxxm-t/MTP-GGUF qwen4exp-mtp-head-Q8_0GGUF (MTP draft head only)4.1download————draft_acceptance_pct 75.1%vs BF16L37 @ c7c423f
myllmbox/hibrid47 hibrid47 (NVFP4 PLE table)NVFP4 mixed (vLLM)—————render_gauntlet_good 26 good; render_gauntlet_partial 3 partial; render_gauntlet_broken 3 brokenother or no referenceL30 @ 7b83fa0
myllmbox/hibrid47-uncensored modified weightshibrid47compressed-tensors NVFP4/FP8 (per HF tags)——————no quality claim@ e2edbc4
myllmbox/hibrid48 hibrid48NVFP4108.5resident————HumanEval pass@1 95.7; GSM8K (200) 98.0; IFEval prompt-strict 91.5; IFEval instruction-strict 93.4; MMLU-Pro (200) 84.9other or no referenceL96 @ e316057
myllmbox/hibrid48-uncensored modified weightshibrid48-uncensorednot stated——————no quality claim@ 572e216
Myric/APEX-GGUF BF16GGUF354.0download—————other or no referenceL31 @ daaeccd
Myric/APEX-GGUF MIDDLEGGUF91.70download————agentic:tasks 9 of 9; agentic:tests 154 of 154; vision:with image 5 of 5other or no referenceL30 @ daaeccd
nazeshinjite/Q8-UDQ4K-DS4 Q8-UDQ4KGGUF (ds4)202.5download————nll 0.2842; api_top1 92.96%; topn_recall 89.49%; pair_order_agree 92.13%; top5_mae 0.0917; greedy_first_token 85/100; nll:long 0.1207; dnll_vs_base −0.0064other or no referenceL94 @ 9374dd4
nibor1896/CNQ4.5-M CNQ4.5-MCNQ (NVFP4)104.7download—82.4 %——oracle_kld (298 rows) 0.343; oracle_kld (607 rows) 0.461; ten tasks greedy 2 Pass / 5 Partial / 3 Fail of 10other or no referenceL418 @ ce0ddcc
Nielk38/GGUF-SPLIT UD-Q2_K_XL (unsloth, re-split)GGUF78.87download—————no quality claimL27 @ b816449
nirecom/mixed-3_8bit mixed-3_8bitMLX85.00download————IFEval prompt_strict (thinking off) 0.8447; IFEval prompt_loose (thinking off) 0.8743; IFEval inst_strict (thinking off) 0.8969; IFEval inst_loose (thinking off) 0.9185; IFEval prompt_strict (thinking on) 0.9279; IFEval prompt_loose (thinking on) 0.9464; IFEval inst_strict (thinking on) 0.9484; IFEval inst_loose (thinking on) 0.9628other or no referenceL89 @ a158876
nitinpanj/Q4_0-Q8out-v3-GGUF Q4_0-Q8out v3GGUF103.1download—————no quality claimL40 @ 967e141
nitinpanj/qwen38-flash-next-v3 Q4_0-Q8out-v3GGUF102.5download——4.3148−17.8%draft_acceptance 0.817other or no referenceL163 @ e298205
nmatteo3294/Q3_PLE-MTP-GGUF AD-4.27bpw-Q3_PLE-M64 targetGGUF (private Q3_PLE type, patched llama.cpp)78.53download—————no quality claimL36 @ 6650250
nmatteo3294/Q3_PLE-MTP-GGUF MTP sidecar DOWNQ4-FC-HC-OUTQ4GGUF (private Q3_PLE type, patched llama.cpp)2.2download—————no quality claimL38 @ 6650250
nopmobiel/MLX-4bit MLX 4-bit g64MLX——————no quality claimL89 @ 8577fa9
nota-ai/Nota-NVFP4 NVFP4NVFP4186.4download————GPQA-Diamond 90.40; HumanEval 98.17; AIME 2026 93.33; HRM8K (KSM) 78.64; BFCL v4 78.83vs BF16L67 @ 53f1bea
nvidia/FP4-Local modified weightsFP4-LocalNVFP4 (+MXFP8)103.1download————GPQA Diamond 91.5%; ToolEvalBench 88.7%vs BF16L118 @ b9ce18a
nvidia/NVFP4 NVFP4NVFP4132.7download (HF listing)————GPQA Diamond 91.5; HLE 35.4; τ²-Bench Telecom 90.1; MMMU Pro 78.3; SciCode 18.8; AA-LCR 74.1; IFBench 81.0; Omniscience 27.6; Terminal-Bench 2.1 82.9other or no referenceL89 @ fc694b5
octolix/Octojet-NVFP4 Octojet NVFP4 (production build)NVFP4 experts + MLX affine 4-bit (Octojet-only)—————GSM8K (250) 245; HumanEval pass@1 (164) 155other or no referenceL105 @ a713d8c
orcarouter/Uncensored-FP8 modified weightsUncensored FP8FP8 (per repo name)——————no quality claim@ a920884
orcarouter/Uncensored-GGUF modified weights(no variants listed)GGUF——————no quality claim@ d27c0a1
orcarouter/Uncensored-MLX modified weightsUncensored MLXMLX (per repo name)——————no quality claim@ fb8c6ba
orcarouter/Uncensored-NVFP4 modified weightsUncensored NVFP4compressed-tensors NVFP4/FP8 (per HF tags)——————no quality claim@ cd56f6c
OsaurusAI/Allosaurus-v0.1-125B-A6B-JANGH2 Allosaurus JANGH2MLX (JANGH)37.37resident————median_kl:general 0.0471; mean_kl:general 0.425; top1:general 81.32%; median_kl:reasoning 0.0086; mean_kl:reasoning 0.108; top1:reasoning 91.58%; median_kl:tools 0.0024; mean_kl:tools 0.677; top1:tools 83.31%; median_kl:office 0.0053; mean_kl:office 0.107; top1:office 91.63%vs BF16L72 @ 01d629e
otheru/Strix-Halo-GGUF modified weightsabliterated native .hgn (FP8 n-gram table)GGUF + halogen .hgn124.4download————ngram_table_rel_rmse:FP8 2.6551%other or no referenceL42 @ 86d977f
otheru/Strix-Halo-GGUF modified weightsabliterated IQ4_XS GGUFGGUF + halogen .hgn97.62download————ngram_table_rel_rmse:IQ4_NL 7.6200%other or no referenceL41 @ 86d977f
otheru/Strix-Halo-GGUF modified weightsqwen38-flash-next-mtp.hgnGGUF + halogen .hgn1.5download—————other or no referenceL43 @ 86d977f
p4ik/MLX-BEAM-4bit BEAM-4bit (pipenetwork mixed-4_8bit repack)MLX (B.E.A.M. layout)106.0download—————no quality claimL23 @ 29e259b
PaoAI/PaoAI-STRIX-BALANCED-2-GGUF BALANCED-2 (previous)GGUF80.78download——4.6695—chain test median 86.7other or no referenceL59 @ 862ba28
PaoAI/PaoAI-STRIX-BALANCED-2-GGUF BALANCED-2.1GGUF78.27download——4.6618−0.16 %chain test per-task median 100/100; PaoAI Quality Bench v1.3 overall 75.7other or no referenceL163 @ 862ba28
paperniuk/MTP-block MTP-blockGGUF (MTP block only)1.5download—————no quality claimL24 @ d98c24a
peasantsmith/PS-IQ2_XXS-GGUF PS-IQ2_XXSGGUF75.20download—————no quality claimL19 @ c8122ff
pentacoxian-dev/IQ3E-Q8D-MTP-GGUF IQ3E-Q8D-MTPGGUF85.92download——2.8506–2.8520—ppl:layer_split 2.8525other or no referenceL30 @ 75e6411
peonist-ai/halogen- w4b (0.14)halogen .hgn (custom)124.1download—————no quality claimL168 @ 43898d3
peonist-ai/halogen- v2 (default, 0.15+)halogen .hgn (custom)66.68trunk only—————no quality claimL158 @ 43898d3
peonist-ai/halogen- ht43 (opt-in)halogen .hgn (custom)57.66trunk only—————no quality claimL52 @ 43898d3
perkel/MC Flash-Next-MCSmallMegaCapybara (repacked EXL3)80.00download—91.7% top1:held-out——kl:held-out 0.0613; kl:own chat/tools 0.0283; top1:own chat/tools 93.7%; kl:agentic 0.1260; top1:agentic 88.7%vs BF16L66 @ 1940705
perkel/MC Flash-Next-MCTinyMegaCapybara (repacked EXL3)65.00download—86.7% top1:held-out——kl:held-out 0.1454; kl:own chat/tools 0.0760; top1:own chat/tools 90.0%; kl:agentic 0.2234; top1:agentic 85.2%vs BF16L66 @ 1940705
petergilani/MLX-6bit 6-bit (n-gram from disk)MLX106.5trunk only——4.4765——reference not statedL46 @ ce5f824
pfeifferj/GSQ-RCO-GGUF GSQ-RCO 3.5bit main + Q4_0/Q8_0 embedding shardGGUF77.42download—————vs BF16L120 @ df4f5bb
pfeifferj/GSQ-RCO-GGUF GSQ-RCO 3.5bit main + BF16 embedding shardGGUF47.94trunk only0.111225 approx_kld—3.1058—MMLU-Pro 58.25%; MMLU-Pro SE 1.10 pp; IFEval strict 13/16; IFEval completed correct 12/16; GSM8K completed correct 8/8vs BF16L82 @ df4f5bb
pipenetwork/MLX-4bit 4bitMLX103.8download——5.3914+20.6%dnll_vs_bf16 +0.1872 [+0.1778, +0.1968]; windows_worse 145/145vs BF16L83 @ aa7c790
pipenetwork/MLX-6bit 6bitMLX148.0download——4.4767—dnll_vs_bf16 +0.0013 [−0.0003, +0.0029]; windows_worse 81/145vs BF16L81 @ b4ccbdb
pipenetwork/MLX-8bit 8bitMLX192.2download——4.4749—dnll_vs_bf16 +0.0009 [−0.0003, +0.0021]; windows_worse 73/145vs BF16L80 @ f1c23d4
pipenetwork/MLX-mixed-4_8bit mixed-4_8bitMLX106.2download——4.5286+1.3%dnll_vs_bf16 +0.0128 [+0.0109, +0.0148]; windows_worse 128/145vs BF16L82 @ b2c422f
Pragmatism0220/Huihui-v2-GSQ-RCO-Strata-abliterated-direction-GGUF modified weightsHuihui-Qwen3.8-Flash-Next-refusal-direction-r-v2.ggufGGUF control vector0.00048download—————no quality claimL31 @ 8003251
primitive-ai/NVFP4 NVFP4NVFP4186.0download————overall 90.2; knowledge 92.2; tool_call 84.6; tool_abstain 56.7; finished_pct 99.4%; GSM8K 98.0; MMLU-Pro 90.0other or no referenceL285 @ e1b0da6
primitive-ai/mixed-NVFP4-FP8 mixed NVFP4/FP8 (v2)NVFP4+FP8183.7download————overall 90.3; knowledge 92.2; tool_call 84.8; tool_abstain 56.7; finished_pct 99.5%other or no referenceL322 @ 07915ee
pugant/MTP-DRAFT-HEAD-ROCMFP4-GGUF Qwen3.8-Flash-Next-MTP-Q8_0-HEAD-ROCMFP4.ggufGGUF (MTP draft head only; ROCmFP4 output tensor)3.8download—————no quality claimL85 @ ba3fe94
pugant/ROCMFP4_STRIX_LEAN-GGUF ROCmFP4 STRIX_LEANGGUF105.8download————ppl:Italian holdout (Dante) 1.1776; ppl:EN calibration 3.1961other or no referenceL53 @ a41ebff
pugant/ROCMFP4_STRIX_LEAN-GGUF ROCmFP2 STRIX_LEAN 2-bit mix (v2)GGUF81.46download————ppl:Italian holdout (Dante) 2.0676 ± 0.011 (1.76×); ppl:EN calibration 3.2689 ± 0.015 (1.02×)other or no referenceL54 @ a41ebff
qtum/GGUF IQ4_XSGGUF127.6download——1.8378 ± 0.05554——vs BF16L40 @ ee96c02
qtum/GGUF IQ3_XXSGGUF115.3download——2.0528 ± 0.06396——vs BF16L41 @ ee96c02
qtum/GGUF IQ2_XS-layeredGGUF89.66download——2.7833 ± 0.10058——vs BF16L42 @ ee96c02
qtum/GGUF IQ1_M-layeredGGUF82.03download——4.4105 ± 0.19322——vs BF16L43 @ ee96c02
Quat3rnion/Signal-3.8-Flash-Next-HGN-v2 modified weightsHGN v2 (assembled 14-file package)HGN (Halogen)118.8download—————no quality claimL90 @ d93045e
Quat3rnion/halogen-v2-abliterated modified weightsnative v2 (current)HGN (Halogen)66.69download—————other or no referenceL48 @ 2fd9605
Quat3rnion/halogen-v2-abliterated modified weightsHT43 abliterated (experimental)HGN (Halogen)57.62download—————other or no referenceL278 @ 2fd9605
Quat3rnion/halogen-swift1.5-v2 modified weightsv2 (swift15)HGN (Halogen)66.69trunk only—————no quality claimL40 @ d6faf17
Quat3rnion/halogen-swift1.5-v2 modified weightsHT43 (swift15)HGN (Halogen)57.62trunk only—————no quality claimL41 @ d6faf17
quimmedes/MTP-GGUF mtp Q4_0GGUF (MTP draft head only)2.2download—————no quality claimL69 @ e56ac7b
quimmedes/MTP-GGUF mtp Q3_0GGUF (MTP draft head only)1.5download—————no quality claimL68 @ e56ac7b
quimmedes/MTP-GGUF mtp Q2_0GGUF (MTP draft head only)1.1download—————no quality claimL67 @ e56ac7b
Qwen/FP8 FP8FP8——————no quality claim@ 236dfdf
r0b0tlab/EXL3-2.50bpw EXL3 2.50bpwEXL365.50download————gsm8k 79/80 — 98.75 %; hard_reasoning 20/20; humaneval 40/40 — 100 %; ifeval 34/39 — 87.18 %; overall 173 correct / 6 incorrect / 1 ungradedother or no referenceL129 @ 61a1a13
r0b0tlab/NVFP4-W4A16-sm121 NVFP4 W4A16NVFP4 W4A16 (ModelOpt)—————Q200-v2 overall 188/200 (94.0%); GSM8K (80) 80; HumanEval (40) 39; IFEval (40) 36; Hard reasoning (20) 20; BFCL v4 structural-hard20 13other or no referenceL68 @ 8b73dc6
RadixArk/NVFP4 NVFP4NVFP4135.0download————GSM8K 97.27; AIME26 pass@1 98.75vs BF16L120 @ 7b71922
rapid-mlx/4bit 4bitMLX105.0download————MMLU-Redux 2.0 86.8 ± 2.1; HumanEval pass@1 96.0 ± 2.0; GSM8K flexible-extract 81.0 ± 3.9; GSM8K answer-aware re-score 96.0; IFEval prompt strict/loose 84.0 / 88.0; IFEval inst strict/loose 89.0 / 92.6other or no referenceL174 @ afeb96d
RedHatAI/NVFP4 NVFP4 (RedHatAI)NVFP4 (compressed-tensors)—————GPQA Diamond 92.9; MMLU-Pro 88.2; AIME 2026 100; LiveCodeBench 95.0; Terminal-Bench 2.1 86.1; SWE-bench Verified 79.4; SWE-bench Pro 63.6; DeepSWE 1.1 62.3; recovery:MMLU-Pro 99.6; recovery:SWE-bench Verified 98.5; recovery:DeepSWE 1.1 94.3; recovery_avg 99.1vs BF16L77 @ c8f2fb1
RESMP-DEV/IQ2_XS-Q2_0-Standard-iMatrix-GGUF IQ2_XS / Q2_0 standard-iMatrixGGUF66.65download—————no quality claimL61 @ 694d57e
risk-magician/AMX_Q8R.amxq AMX_Q8RAMX_Q8R (.amxq)——————no quality claim@ 50647bc
robertbschuette/Huihui-abliterated-GGUF modified weightsabliterated GGUF series (unsloth / Swift)GGUF——————no quality claim@ 24b9683
Robot-Haus/MLX-oQ3-MTP oQ3-MTPMLX92.50download————MMLU 80.0%; TruthfulQA 100.0%; HumanEval 90.0%; MBPP 90.0%other or no referenceL76 @ 9ad88db
Robot-Haus/oQ3.5e-fp16-mtp oQ3.5e-fp16MLX—————MMLU 85.4%; HellaSwag 94.3%; TruthfulQA 95.3%; HumanEval pass@1 91.5%; MBPP pass@1 87.5%; LiveCodeBench pass@1 48.0%; SafetyBench 86.7%other or no referenceL171 @ 612c2b7
Robot-Haus/oQ4e-fp16-mtp oQ4e fp16MLX107.2download————agentic_battery 3/3other or no referenceL22 @ c220aa3
RockTalk/MLX-4bit MLX 4-bit g64MLX71.00trunk only—————no quality claimL172 @ 0afead4
SargeDev/lil-rascal-uncensored modified weightsNVFP4/MXFP8 uncensoredNVFP4/MXFP8——————no quality claim@ e20e53c
Sawfwair/MLX-4bit MLX 4-bitMLX104.7download—————no quality claimL22 @ 6cc9bbc
Sawfwair/MLX-Activation-3bit Activation-weighted 3-bitMLX89.65download————suite_exact_pass 58 of 61; holdout_identical_to_Q4 all 16other or no referenceL67 @ c699bd6
Sawfwair/MLX-Activation-3bit-Native-PLE Activation-weighted 3-bit Native PLEMLX89.65download————cases_exact 58 of 61; holdout_identical_to_Q4 all 16other or no referenceL29 @ 1cee930
Sawfwair/MLX-Mixed-2bit Mixed Q2/Q4MLX73.10download—————no quality claimL22 @ a6e3d7a
scorpoon/exl3-SC 3.15bpw h8 SCEXL354.16GPU weights0.01454—1.4016+1.23%—vs BF16L61 @ 8165296
scorpoon/exl3-SC 3.05bpw h8 SCEXL353.21GPU weights0.01543—1.4041+1.41%—vs BF16L60 @ 8165296
scorpoon/exl3-SC 2.05bpw h8 SCEXL336.78GPU weights0.05761—1.4697+6.15%—vs BF16L62 @ 8165296
scottlowry/oQ3.5e-fp16-mtp oQ3.5e-fp16MLX——————no quality claim@ d203e0a
scottlowry/oQ3.5e-mtp oQ3.5eMLX——————no quality claim@ 2c1f59f
scottlowry/oQ4e-fp16-mtp oQ4e-fp16MLX——————no quality claim@ aa8cb7c
scottlowry/oQ4e-mtp oQ4eMLX——————no quality claim@ 998d2d5
scottlowry/oQ5e-fp16-mtp oQ5e-fp16MLX——————no quality claim@ f876fb2
scottlowry/oQ5e-mtp oQ5eMLX——————no quality claim@ 38b03ce
scottlowry/oQ6e-fp16-mtp oQ6e-fp16MLX——————no quality claim@ 0d81135
scottlowry/oQ6e-mtp oQ6eMLX——————no quality claim@ 9cb47a7
scottlowry/oQ8e-fp16-mtp oQ8e-fp16MLX——————no quality claim@ 6e1875b
scottlowry/oQ8e-mtp oQ8eMLX——————no quality claim@ 4bf7525
ScottzillaSystems/CYBERSECURITY-NVFP4 modified weightsCYBERSECURITY NVFP4NVFP4—————MMLU 81.27%; MMLU delta -0.84 pp; hard multi-step accuracy 98.3%other or no referenceL134 @ b62c93b
scouzi1966/AFM-MLX-4bit afm-mapped-4MLX (AFM)——————no quality claim@ 3114e33
sh0wie/REAP-288-GGUF modified weightsQ8_0GGUF116.0download—————other or no referenceL45 @ 2af8d94
sh0wie/REAP-288-GGUF modified weightsQ5_K_MGGUF87.00download—————other or no referenceL44 @ 2af8d94
sh0wie/REAP-288-GGUF modified weightsQ4_K_MGGUF78.00download—————other or no referenceL43 @ 2af8d94
sh0wie/REAP-288-MLX-4bit modified weightsREAP-288 4-bitMLX68.00download————HumanEval pass@1 91.5%; rare_name_sampling 9 of 10other or no referenceL115 @ 668f31b
sh0wie/REAP-288-Q8E-MLX modified weightsREAP-288 Q8EMLX100.0download————HumanEval pass@1 90.9%; rare_name_reliability 9 of 10other or no referenceL21 @ e3e7da4
sh0wie/REAP-384-MLX-4bit modified weightsREAP-384 4-bitMLX80.00download————HumanEval pass@1 92.1%other or no referenceL21 @ c9bb01b
showkieo/Sushi-3bpw Sushi-3bpw (4-bit n-gram table)EXL3 (sushi pack)52.97GPU weights0.104790.34%———vs BF16L118 @ 851e86b
SiliconSpecies/EXL3-Sushi 6bpw-hqEXL3 (Sushi pack)98.78trunk only—93.5%——kld:wikitext 0.0284; kld:reasoning 0.0099; kld:32K 0.0124vs BF16L36 @ 950fcc1
SiliconSpecies/EXL3-Sushi 5bpwEXL3 (Sushi pack)83.75trunk only—92.9%——kld:wikitext 0.0357; kld:reasoning 0.0105; kld:32K 0.0163vs BF16L34 @ 950fcc1
SiliconSpecies/EXL3-Sushi 5bpw-hqEXL3 (Sushi pack)83.75trunk only—93.1%——kld:wikitext 0.0325; kld:reasoning 0.0095; kld:32K 0.0153vs BF16L35 @ 950fcc1
sirfyyn/FLE-GGUF FLE-IQ3_MGGUF88.87download————kld:E 0.051 ± 0.003; kld:W 0.050 ± 0.001; kld:K 0.129 ± 0.004; top1_pct:E/W/K 96.9 / 91.5 / 89.1; IFBench prompt strict 80.3; GPQA Diamond 88.9; LiveCodeBench v6 85.5vs BF16L218 @ 1b64f8e
sirfyyn/FLE-GGUF FLE-IQ3_XXSGGUF79.48download————kld:E 0.090 ± 0.005; kld:W 0.093 ± 0.002; kld:K 0.227 ± 0.007; top1_pct:E/W/K 95.7 / 88.1 / 84.2vs BF16L224 @ 1b64f8e
sirfyyn/FLE-GGUF FLE-IQ2_MGGUF71.62download————kld:E 0.160 ± 0.007; kld:W 0.163 ± 0.004; kld:K 0.339 ± 0.008; top1_pct:E/W/K 94.0 / 84.7 / 80.9vs BF16L228 @ 1b64f8e
sirfyyn/FLE-REAP448-GGUF modified weightsFLE-IQ3_M-REAP448GGUF82.00download————kld:E 0.080 ± 0.005; kld:W 0.086 ± 0.003; kld:K 0.185 ± 0.006; top1_pct E/W/K 96.0 / 89.7 / 87.2; MKQA German 21.5 %; MKQA English 29.8 %vs BF16L87 @ 6bbaae7
sleepyeldrazi/GSQ-RCO-IQ3_XXS-HCQ5_0 IQ3_XXS-HCQ5_0GGUF49.61trunk only————greedy_identical_prefix_tokens 47reference not statedL68 @ cb186e7
Slomin/Potato-CODER-48GB modified weightsPotato-CODER-48GB IQ4_XS (v1.0, withdrawn)GGUF65.49download————lang:bash 98.8; lang:c 99.6; lang:cpp 92.5; lang:csharp 99.6; lang:dart 96.7; lang:flutter (widget tests) 94.6; lang:go 99.6; lang:java 100.0; lang:javascript 96.7; lang:kotlin 96.2; lang:php 87.1; lang:python 92.9; lang:ruby 94.6; lang:rust 95.4; lang:swift 94.2; lang:typescript 95.0; dev:LiveCodeBench (144) 125; dev:16 programming languages (640 tasks) 616; dev:8 human languages (160) 157; dev:tool-call suite (18) 17; dev:HumanEval+ / MBPP+ (264) 229; dev:BFCL v3 (200) 148other or no referenceL91 @ a96007b
sorenmat/NVFP4 NVFP4NVFP4—————GPQA Diamond 91.5; HLE 35.4; τ²-Bench Telecom 90.1; MMMU Pro 78.3; SciCode 18.8; AA-LCR 74.1; IFBench 81.0; Omniscience 27.6; Terminal-Bench 2.1 82.9other or no referenceL91 @ 7023287
spectator2026/W6-HQQ-g32 W6 HQQ g32compressed-tensors int6 (W6A16)224.0download————kld:wikitext fp8KV 0.0228; kld:wikitext bf16KV 0.0210; kld:agentic fp8KV 0.0749; kld:agentic bf16KV 0.0672; kld:wikitext pinned image 0.0223vs BF16L110 @ 05848f8
SpeederX/UD-Q2_K_XL-EM-GGUF UD-Q2_K_XL uniform IQ2_XS expert-majorGGUF78.88download————top1_agreement 61 of 65; mean_top10_kl 0.095; max_top10_kl 0.76other or no referenceL130 @ 2a6e332
spiritfather/cyberneurova-GGUF modified weightsQ8_0GGUF188.2download—————no quality claimL47 @ cd72aeb
spiritfather/cyberneurova-GGUF modified weightsQ6_KGGUF167.6download—————no quality claimL46 @ cd72aeb
spiritfather/cyberneurova-GGUF modified weightsQ5_K_MGGUF134.1download—————no quality claimL45 @ cd72aeb
spiritfather/cyberneurova-GGUF modified weightsQ5_K_SGGUF127.8download—————no quality claimL44 @ cd72aeb
spiritfather/cyberneurova-GGUF modified weightsQ4_K_MGGUF119.2download—————no quality claimL43 @ cd72aeb
spiritfather/cyberneurova-GGUF modified weightsQ4_K_SGGUF111.8download—————no quality claimL42 @ cd72aeb
spiritfather/cyberneurova-GGUF modified weightsQ3_K_LGGUF96.60download—————no quality claimL41 @ cd72aeb
spiritfather/cyberneurova-GGUF modified weightsQ3_K_MGGUF94.20download—————no quality claimL40 @ cd72aeb
spiritfather/cyberneurova-GGUF modified weightsQ3_K_SGGUF88.80download—————no quality claimL39 @ cd72aeb
spiritfather/cyberneurova-GGUF modified weightsQ2_KGGUF80.40download—————no quality claimL38 @ cd72aeb
spiritfather/cyberneurova-i1-GGUF modified weightsi1-Q6_KGGUF167.6download—————other or no referenceL71 @ d2d2635
spiritfather/cyberneurova-i1-GGUF modified weightsi1-Q5_K_MGGUF134.1download—————other or no referenceL70 @ d2d2635
spiritfather/cyberneurova-i1-GGUF modified weightsi1-Q5_K_SGGUF127.8download—————other or no referenceL69 @ d2d2635
spiritfather/cyberneurova-i1-GGUF modified weightsIQ4_XS (PLE mixed)GGUF125.3download0.039888 ± 0.000269—4.8502 ± 0.030+0.28%—other or no referenceL51 @ d2d2635
spiritfather/cyberneurova-i1-GGUF modified weightsi1-Q4_K_MGGUF119.2download—————other or no referenceL67 @ d2d2635
spiritfather/cyberneurova-i1-GGUF modified weightsIQ3_M (PLE mixed)GGUF117.2download0.072916 ± 0.000464—4.9869 ± 0.031+3.11%—other or no referenceL50 @ d2d2635
spiritfather/cyberneurova-i1-GGUF modified weightsIQ3_XXS (PLE mixed)GGUF113.4download0.093681 ± 0.000580—4.9843 ± 0.031+3.05%—other or no referenceL49 @ d2d2635
spiritfather/cyberneurova-i1-GGUF modified weightsi1-Q4_1GGUF111.0download—————other or no referenceL68 @ d2d2635
spiritfather/cyberneurova-i1-GGUF modified weightsIQ2_M (PLE mixed)GGUF108.3download0.150323 ± 0.000859—5.1252 ± 0.032+5.97%—other or no referenceL48 @ d2d2635
spiritfather/cyberneurova-i1-GGUF modified weightsi1-Q4_K_SGGUF106.2download—————other or no referenceL66 @ d2d2635
spiritfather/cyberneurova-i1-GGUF modified weightsi1-Q4_0GGUF100.3download—————other or no referenceL65 @ d2d2635
spiritfather/cyberneurova-i1-GGUF modified weightsi1-IQ4_NLGGUF100.1download—————other or no referenceL64 @ d2d2635
spiritfather/cyberneurova-i1-GGUF modified weightsi1-Q3_K_LGGUF96.60download—————other or no referenceL63 @ d2d2635
spiritfather/cyberneurova-i1-GGUF modified weightsi1-Q3_K_MGGUF94.20download—————other or no referenceL62 @ d2d2635
spiritfather/cyberneurova-i1-GGUF modified weightsi1-IQ3_SGGUF88.90download—————other or no referenceL60 @ d2d2635
spiritfather/cyberneurova-i1-GGUF modified weightsi1-Q3_K_SGGUF88.80download—————other or no referenceL61 @ d2d2635
spiritfather/cyberneurova-i1-GGUF modified weightsi1-Q2_KGGUF80.40download—————other or no referenceL59 @ d2d2635
starkweatherdigital/nvfp4 NVFP4NVFP4 (ModelOpt)109.0download—————no quality claimL18 @ 1b304e5
steadfastgaze/MoEspressoV3 IQ_K routed experts, Cache-Prior 2/2, capacity 223, mediumMoEspresso package (IQ_K/K-quant MLX)145.8download————LiveBench 48q six-category macro 84.3%other or no referenceL89 @ 7c83023
stgallenquants/GGUF GGUF (files not listed on card)GGUF——————no quality claim@ 34cf440
SUNNYROAM/Huihui-abliterated-GGUF modified weightsUD-Q4_K_XL (unsloth series)GGUF——————no quality claim@ 3110adf
SUNNYROAM/Huihui-abliterated-GGUF modified weightsSwift seriesGGUF——————no quality claim@ 3110adf
SuperPUPS/INT8x17-AttnINT4 INT8x17-AttnINT4compressed-tensors (INT8/INT4 pack-quantized)202.4download————int8_roundtrip_cos_vs_fp8 0.999962; golden_probes 3/3; 10-task battery 10/10; determinism_8x 7/8; niah_128k HIT; soak_errors 0 errorsother or no referenceL33 @ 8d7e6f7
SuperPUPS/INT8x25-AttnBF16 INT8x25-AttnBF16compressed-tensors (INT8/INT4 experts, BF16 attention)212.0download————int8_roundtrip_cos 0.99995; int8_roundtrip_cos_worst 0.99990; golden_probes 3/3; battery_10task 10/10; niah_128k HITother or no referenceL96 @ 6c077c3
SZYBKIMARCIN955/GGUF not listedGGUF——————no quality claim@ 5ef30cd
talhasarac/REAP128-86B-Unity-OpenCode-GGUF modified weightsREAP128 UD-IQ4_XSGGUF48.85download—————no quality claimL58 @ d396d76
tapsin/TAPSIN-Qwen3.8-4B-Distill-GGUF modified weightsBF16GGUF8.7download—————no quality claimL36 @ 994c589
tapsin/TAPSIN-Qwen3.8-4B-Distill-GGUF modified weightsQ8_0GGUF4.6download—————no quality claimL35 @ 994c589
tapsin/TAPSIN-Qwen3.8-4B-Distill-GGUF modified weightsQ6_KGGUF3.6download—————no quality claimL34 @ 994c589
tapsin/TAPSIN-Qwen3.8-4B-Distill-GGUF modified weightsQ5_K_MGGUF3.2download—————no quality claimL33 @ 994c589
tapsin/TAPSIN-Qwen3.8-4B-Distill-GGUF modified weightsQ4_K_MGGUF2.8download—————no quality claimL32 @ 994c589
tarruda/GGUF Q5_KGGUF——————no quality claimL26 @ 0eb873e
tarruda/GGUF IQ4_NLGGUF——————no quality claimL27 @ 0eb873e
tawei-amd/strixflash-GGUF strixflash-mixAGGUF——————no quality claim@ 2cdd89a
tcclaviger/MXFP4-FP8 MXFP4/FP8MXFP4/FP8125.8download—————no quality claimL19 @ 58ccaf6
tcclaviger/MXFP4-FP8-GPTQ MXFP4-FP8 AA2MXFP4+FP8116.0download.054995.17%4.552——vs BF16L24 @ 6fe3e74
TensorFold/MLX-2bit-MTP MLX 2-bit g32 + BF16 exclusionsMLX80.07download—————no quality claimL78 @ 28b1090
TensorFold/MLX-3bit-MTP MLX 3-bit g32MLX90.77download—————no quality claimL78 @ 1d53b5c
TensorFold/MLX-4bit MLX 4-bit g32MLX111.6download—————no quality claimL64 @ 4fc403b
TensorFold/MLX-4bit-MTP MLX 4-bit g32 MTPMLX113.2download—————no quality claimL78 @ 2b170fa
TensorFold/MLX-6bit-MTP MLX 6-bit g32MLX158.1download—————no quality claimL78 @ 45032a1
TensorFold/MLX-8bit-MTP MLX 8-bit g32MLX203.0download—————no quality claimL78 @ a98d2ff
TensorFold/MLX-oQ2 oQ2MLX67.67download—————no quality claimL69 @ 75e9a28
TensorFold/MLX-oQ2-MTP oQ2MLX70.78download—————no quality claimL80 @ 0a1bf6f
TensorFold/MLX-oQ3-MTP oQ3MLX92.50download—————no quality claimL79 @ 993186e
TensorFold/MLX-oQ4 oQ4MLX111.7download—————no quality claimL65 @ dde9d8d
TensorFold/MLX-oQ4-MTP oQ4 MTPMLX113.3download—————no quality claimL77 @ 069090c
TensorFold/MLX-oQ6 oQ6MLX155.8download—————no quality claimL65 @ aabd8cd
TensorFold/MLX-oQ6-MTP oQ6MLX (oQ)158.1download—————no quality claimL76 @ d6209ba
TensorFold/MLX-oQ8-MTP MLX oQ8MLX202.8download—————no quality claimL76 @ eb86370
TheDrainFlorist/VQ-2.1bpw VQ-2.1bpwMLX47.78text weights (excl. vision, MTP)652.4 mnats72.3%3.559——vs BF16L173 @ b5591de
TheDrainFlorist/VQ-3.2bpw VQ-3.2bpwMLX73.77text weights (excl. vision, MTP)147.3 mnats86.7%2.300—kl_mnats:prose 305.6; kl_mnats:code 81.8; kl_mnats:literary 54.6vs BF16L173 @ 33e10e6
TheDrainFlorist/VQ-4.4bpw VQ-4.4bpwMLX100.5text weights (excl. vision, MTP)47.7 mnats92.2%2.323——vs BF16L184 @ c13ef42
TheDrainFlorist/VQ-5.5bpw VQ-5.5bpwMLX119.8text weights (excl. vision, MTP)30.9 mnats93.7%2.244—kl_mnats:prose 68.8; kl_mnats:code 16.5; kl_mnats:literary 7.3vs BF16L163 @ 7740086
thoddnn/4bit MLX 4-bit g32MLX——————no quality claim@ b070255
thousandml/CQ-4bit CQ 4-bitCQ 4-bit (custom)——————no quality claim@ a4492d6
tirex2001/DACAN Q8_0 expertsDACAN pack (GGUF dense + expert .bin)128.0download—————no quality claimL62 @ bbd468e
tirex2001/DACAN NVFP4 experts (from NVIDIA)DACAN pack (GGUF dense + expert .bin)68.00download—————no quality claimL63 @ bbd468e
tirex2001/DACAN PLE FP8->Q8_0 GGUFDACAN pack (GGUF dense + expert .bin)54.00download—————no quality claimL65 @ bbd468e
tirex2001/DACAN dense Q8_0 GGUFDACAN pack (GGUF dense + expert .bin)6download—————no quality claimL64 @ bbd468e
tiyuvta/NVFP4 NVFP4 + BF16 MTPNVFP4 (ModelOpt)—————kl:merge two sorted linked lists 0.0272; top20_overlap:merge two sorted linked lists 18/20; argmax_same:merge two sorted linked lists yes; kl:TCP vs UDP in two sentences 0.0035; top20_overlap:TCP vs UDP in two sentences 16/20; kl:def fib(n) continuation 0.0124; top20_overlap:def fib(n) continuation 19/20; kl:capital of Australia 0.0355; top20_overlap:capital of Australia 18/20; argmax_same:capital of Australia novs BF16L46 @ 7154444
tls20/oQ5e-mtp oQ5e-mtpMLX (oMLX oQ5e)—————MBPP 88.6%; HumanEval 94.5%; MMLU 92%; JMMLU 90.6%; TruthfulQA 95.3%; LiveCodeBench 59%other or no referenceL83 @ 6bffcfd
ToPo-ToPo/mlx-4bit MLX 4-bit g32MLX111.7download—————no quality claimL18 @ f128069
ToPo-ToPo/mlx-8bit MLX 8-bit g32MLX199.7download—————no quality claimL18 @ ac6a8c8
TracNetwork/mayhem-catalog-Qwen-Qwen3-8-Flash-Next-NVFP4 NVFP4NVFP4135.0download————GSM8K 97.27; AIME26 pass@1 98.75vs BF16L120 @ ae294c1
Trnio/qwen-3.8 FP8FP8——————no quality claim@ d44e7a0
trustos/GSQ-RCO-Q2-MTPLX GSQ-RCO Q2 MTPLXMLX (MTPLX)77.30download—————no quality claimL51 @ 09d7bd1
trymirai/S-experimental S-experimentalnot stated——————no quality claim@ af20c48
trymirai/S-small-experimental S-small-experimentalsafetensors (2-bit, per HF tags)——————no quality claim@ c26c47d
turboderp/exl3 6.05 bpw H6 NG6EXL3139.0download (HF branch listing)0.0031—1.3825——vs BF16L27 @ c1def42
turboderp/exl3 5.05 bpw H6 NG6EXL3123.2download (HF branch listing)0.0040—1.3848——vs BF16L27 @ c1def42
turboderp/exl3 4.05 bpw H6 NG6EXL3107.5download (HF branch listing)0.0067—1.3900——vs BF16L27 @ c1def42
turboderp/exl3 3.05 bpw H5 NG5EXL385.14download (HF branch listing)0.0177—1.4075——vs BF16L27 @ c1def42
turboderp/exl3 2.05 bpw H4 NG4EXL362.82download (HF branch listing)0.0684—1.4879——vs BF16L27 @ c1def42
txgsync/Dynamic-oQ5e-BF16-PLE-mtp Dynamic-oQ5e BF16-PLEMLX187.3download—98.0194%1.28318358—kl_vs_bf16 0.01672752; llm-bench mean quality (xhigh, t1.0, seed 6101) 73.86; llm-bench mean quality (xhigh, t1.0, seed 6102) 86.84; llm-bench mean quality (medium, t1.0, seed 6101) 79.13; llm-bench mean quality (medium, t0.6, seed 6101) 78.84vs BF16L94 @ b62713a
txgsync/MXFP4-MLX MXFP4-MLX mixedMLX99.07download—————no quality claimL40 @ 016f76a
Tylogi/EW-MFQ V4-XSMFQ (.mfq)93.67download0.07524 kld90.36%———reference not statedL85 @ 3cdfba4
Tylogi/EW-MFQ S3-LMFQ (.mfq)89.98download0.08698 kld89.55%———reference not statedL84 @ 3cdfba4
Tylogi/EW-MFQ V3-XXSMFQ (.mfq)81.95download0.12436 kld87.62%———reference not statedL83 @ 3cdfba4
Tylogi/EW-MFQ S2-LMFQ (.mfq)78.86download0.14630 kld86.54%———reference not statedL82 @ 3cdfba4
Tylogi/EW-MFQ V1-MMFQ (.mfq)74.53download0.19689 kld84.72%———reference not statedL81 @ 3cdfba4
Tylogi/EW-MFQ V1-SMFQ (.mfq)72.54download0.22157 kld83.77%———reference not statedL80 @ 3cdfba4
Tylogi/EW-MFQ S4-LMFQ (.mfq)——————reference not stated@ 3cdfba4
Tylogi/EW-MFQ S5-LMFQ (.mfq)——————reference not stated@ 3cdfba4
UltimateIntent/Abliterated-s1.5-GGUF modified weightsQ8_0GGUF188.2download—————no quality claimL39 @ dc11a16
UltimateIntent/Abliterated-s1.5-GGUF modified weightsQ6_KGGUF167.6download—————no quality claimL40 @ dc11a16
UltimateIntent/Abliterated-s1.5-GGUF modified weightsQ5_K_MGGUF134.1download—————no quality claimL41 @ dc11a16
UltimateIntent/Abliterated-s1.5-GGUF modified weightsQ4_K_MGGUF119.1download—————no quality claimL42 @ dc11a16
UltimateIntent/Abliterated-s1.5-GGUF modified weightsIQ4_XSGGUF98.42download—————no quality claimL43 @ dc11a16
umixer/oQ4e-mix6-mtp oQ4e-mix6 v2MLX125.4download————MMLU-Pro 86.3%; MMLU-Pro correct 259/300; MMLU-Pro accuracy on finished 91.4%; MATHQA 92.3%; MATHQA correct 277/300; MATHQA accuracy on finished 96.8%; MBPP 89.3%; MBPP correct 268/300; MBPP accuracy on finished 99.3%other or no referenceL144 @ 73d2cd0
unsloth/FP8 FP8FP8——————no quality claim@ 41cc25f
unsloth/GGUF Q8_0GGUF188.2download0.02657494.122 top1——kld_99.9 0.8118reference not statedunsloth.ai docs
unsloth/GGUF UD-Q6_K_XLGGUF169.2download0.02709194.089 top1——kld_99.9 0.8416reference not statedunsloth.ai docs
unsloth/GGUF UD-Q5_K_XLGGUF158.3download0.03041593.680 top1——kld_99.9 1.0036reference not statedunsloth.ai docs
unsloth/GGUF UD-Q4_K_XLGGUF111.3download0.04689392.255 top1——kld_99.9 1.5468reference not statedunsloth.ai docs
unsloth/GGUF UD-IQ4_XSGGUF93.70download0.08363089.554 top1——kld_99.9 2.3677reference not statedunsloth.ai docs
unsloth/GGUF UD-Q3_K_XLGGUF90.00download0.10650488.315 top1——kld_99.9 3.0538reference not statedunsloth.ai docs
unsloth/GGUF UD-IQ3_XXSGGUF82.00download0.16512085.414 top1——kld_99.9 4.0375reference not statedunsloth.ai docs
unsloth/GGUF UD-Q2_K_XLGGUF78.90download0.22460782.715 top1——kld_99.9 4.9121reference not statedunsloth.ai docs
unsloth/GGUF UD-IQ1_MGGUF74.50download0.31473979.691 top1——kld_99.9 6.1965reference not statedunsloth.ai docs
unsloth/GGUF UD-IQ1_SGGUF72.50download0.39607077.325 top1——kld_99.9 7.2126reference not statedunsloth.ai docs
V1ta11y1/Huihui-abliterated-v2-GGUF modified weightsUD-Q4_K_XLGGUF——————no quality claim@ 09de5fa
vaultai/MLX-oQ4-MTP oQ4 MTPMLX113.3download—————no quality claimL72 @ 6b96f4c
vcruz305/EXL3-SAGE-4.15bpw EXL3 SAGE 4.15bpwEXL3109.0download—————no quality claimL44 @ d5d792a
vcruz305/GGUF Q6_KGGUF115.8trunk only—————no quality claimL42 @ 0c50dde
vcruz305/GGUF Q5_K_MGGUF98.03trunk only—————no quality claimL41 @ 0c50dde
vcruz305/GGUF Q4_K_MGGUF86.11trunk only—————no quality claimL40 @ 0c50dde
vcruz305/GGUF Q3_K_MGGUF67.11trunk only—————no quality claimL39 @ 0c50dde
vcruz305/GGUF Q2_KGGUF53.15trunk only—————no quality claimL38 @ 0c50dde
Vemloir/sc-exl3 4.05bpw_sc_h8_ng8EXL369.00trunk only————kl:Nemotron 0.00796; kl:multilingual 0.00544; kl_delta:Nemotron -12.4% [-18.6%, -6.2%]; kl_delta:multilingual -16.9% [-24.6%, -10.1%]; ppl:Nemotron 1.9309; ppl:multilingual 2.3452vs BF16L40 @ c9e1197
Vitruvi0/nvfp4-fp8hybrid nvfp4-fp8hybridNVFP4/FP8 (per repo name)——————no quality claim@ 9a463b4
vkshdev/FP8-180B FP8FP8——————no quality claim@ 1e0c4cf
VnimanieAI/W4A16 INT4 W4A16 g128compressed-tensors W4A16168.0download————greedy suite (TP8+EP) 48/50; greedy suite (TP4+EP) 47/50; τ²-Bench Airline pass^1 0.740; MMLU_RU strict 0.815 ± 0.012; MMLU_RU budget-corrected 0.877 ± 0.010; GPQA Diamond strict @32k 0.601 ± 0.035; GPQA Diamond budget-corrected @131k 0.874 ± 0.024; GPQA paired 40 correct 19/40other or no referenceL102 @ 142456d
vumpt/GGUF Q4_K_MGGUF120.0download—————no quality claimL35 @ 83a79fc
WarlaxZ/NInfer-3090 IQ2_XS (.ninfer)NInfer container (from GGUF)41.00trunk only————ppl:Strata prompts 7.8876; ppl:Infernix prompts 5.0555; dnll:Strata prompts +0.0064 ± 0.0027; dnll:Infernix prompts +0.0007 ± 0.0025; top1:Strata prompts 97.1%; top1:Infernix prompts 97.4%; kl_mean/p99:Strata prompts 0.0042 / 0.039; kl_mean/p99:Infernix prompts 0.0041 / 0.038other or no referenceL130 @ c7dfb96
wemoh/strixw strixw (U-gdn_in-g128)strixw70.76trunk only0.028226 of 242 top1——agent_tasks 19 of 19vs BF16L94 @ b64f3a6
whm0627/177B-A3B-fits64GB-PLElast-GGUF Q2_K_XL-PLElastGGUF79.00download——1.084—GSM8K (40) 37/40; HumanEval (20) 20/20other or no referenceL45 @ 8945f87
whm0627/177B-A3B-fits64GB-PLElast-GGUF IQ1_S-PLElastGGUF68.00download——1.156—GSM8K (40) 38/40; HumanEval (20) 19/20other or no referenceL45 @ 8945f87
Wladastic/Ombro-v13b-GGUF Ombro V13bGGUF118.9download—————no quality claimL27 @ 4b20669
Wladastic/Ombro-v9-GGUF Ombro V9GGUF114.5download—————no quality claimL43 @ ba82efc
wrldsuksgo2mars/EXL3-K4.25-v1 EXL3 K4.25 v1EXL3 (GPTQModel fork)179.4download————ppl:RTX PRO 6000 Blackwell 3.87893; ppl:DGX Spark / GB10 3.85601vs BF16L56 @ 73a050c
wtdcode/AWQ-W4A16 AWQ-W4A16compressed-tensors AWQ INT4 (W4A16)——————no quality claim@ 0939125
xautonomics/UD-IQ4_XS.LAYR.GGUF UD-IQ4_XS (unsloth, sliced)GGUF (per-layer library)94.27download—————no quality claimL25 @ 9ee54f3
xpashkin/MLX-Serve-iQ-MLX-6.6bpw iQ-MLX 6.6 bpwMLX (mlx-serve pack)164.3download—————no quality claimL36 @ fb2a8be
xpashkin/MLX-Serve-iQ-MLX-7.1bpw iQ-MLX 7.1bpwMLX (mlx-serve)172.8download—————no quality claimL62 @ f8e2927
xrrand/MLX-Serve-4bit-mtp mlx-serve 4-bit MTP hybridMLX (mlx-serve layout)—————Tool-Eval-Bench score 90-91; Tool-Eval-Bench hard mode 36/46 (78%); Tool-Eval-Bench points (single run) 167/184other or no referenceL47 @ afd39de
xv0y5ncu/GLQ-trellis-3inst-3bpw-ple4 GLQ 3bpw ple4GLQ83.24download——4.8898—AIME-2026 avg@8 99.58%; AIME-2026 samples 239 of 240other or no referenceL193 @ cc48093
yamz-labs/EXL3-Yamz EXL3 packEXL395.00download0.0216 kl93.81 %——kl_run_to_run 0.0008other or no referenceL83 @ 779f63f
Youssofal/MTPLX-Bare-Speed Bare Speed (flat 4-bit)MLX (MTPLX pack)106.3download—————no quality claimL100 @ 4285a4a
Youssofal/MTPLX-Optimized-Quality Optimized Quality (8-bit)MLX (MTPLX)170.0download—————no quality claimL68 @ 56b2d9f
Youssofal/MTPLX-Optimized-Speed Optimized Speed (4-bit, 8-bit attention)MLX (MTPLX layout)115.1download—————no quality claimL122 @ 34ae535
YTan2000/TQ3_4S TQ3_4SGGUF118.1download————Hard86 72/86 = 83.7%; HumanEval+ base/plus 85.4% / 82.3%; MBPP+ base/plus 89.9% / 76.2%; suite coding 100.0 (12/12); suite data extraction 85.1 (9/15); suite instruction following 83.3 (10/15); suite tool calling 81.7 (12/15); suite reasoning & math 73.3 (11/15)other or no referenceL114 @ 589e20f
Zalypamoy91/Huihui-abliterated-GGUF modified weightsUD-Q4_K_XLGGUF——————no quality claim@ 8894dad
zankich/W4A16-Merlin-INT4PLE-FP8KV W4A16-Merlin + INT4 PLE + FP8 KVcompressed-tensors W4A16 (GPTQ) + INT4 PLE + FP8 KV scales95.00download————MMLU-Pro (700 q, 5-shot) 0.8486 ± 0.0133; MMLU-Pro (112 q subset, FP8 KV) 0.875 ± 0.031other or no referenceL98 @ ba8fff3
zerodigest/REAP-320-YMQ-GGUF modified weightsXS-ProGGUF63.00download—————no quality claimL60 @ 0ac7648
zerodigest/REAP-320-YMQ-GGUF modified weightsM-TIGGUF63.00download—————no quality claimL61 @ 0ac7648
AesSedai/GGUF quoted by agentionai, not measuredQ4_K_MGGUF145.4unknown———+0.61%—vs BF16L172 @ d37a14a
AesSedai/GGUF quoted by agentionai, not measuredIQ4_XSGGUF125.8unknown———+3.12%—vs BF16L171 @ d37a14a
AesSedai/GGUF quoted by agentionai, not measuredIQ3_SGGUF115.3unknown———+6.10%—vs BF16L169 @ d37a14a
agentionai/AP-GGUF measured by agentionaiAP-Q4_K_XLGGUF72.35GPU weights0.10785.52%———vs a Q8_0L65 @ aa0a0c3
agentionai/AP-GGUF measured by jcbtcAP-Q5_K_XLGGUF—0.029135 kl96.436%——tail_ppl 2.02636; hard_tools_per100 73; hermes_pass1_per100 91; hermes_pass2_per100 91vs BF16L190 @ f3f020e
airagrp/mixed-iQ5-Q8-ple16-mlx-serve publisher comparison buildmixed-iQ5-tail8-bf16heads (airagrp repo)MLX—0.068 kld92.4%———vs BF16L117 @ af5fc98
airagrp/mixed-iQ5-Q8-ple16-mlx-serve publisher comparison buildsame widths, no imatrixMLX—0.073 kld91.5%———vs BF16L118 @ af5fc98
airagrp/mixed-iQ5-Q8-ple16-mlx-serve publisher comparison buildsame trunk, 8-bit n-gram tableMLX—0.081 kld89.6%———vs BF16L119 @ af5fc98
airagrp/mixed-iQ5-Q8-ple16-mlx-serve publisher comparison build4-bit-embed mixed pack (local)MLX—0.071 kld92.9%———vs BF16L120 @ af5fc98
aniolekx/E9-MLX-8bit publisher comparison buildE8h predecessor (embed_tokens + n-gram table 4-bit g32)MLX—————nll 1.79914; ppl:technical 6.777; top1:technical 63.18%; nll_delta_vs_near_bf16 0.00116 ± 0.00147other or no referenceL102 @ cfd4a00
aniolekx/E9-MLX-8bit publisher comparison buildnear-bf16 reference buildMLX—————nll 1.80030other or no referenceL102 @ cfd4a00
aniolekx/E9-MLX-8bit publisher comparison buildE9 with 4-bit g32 routed experts (Vontra banks)MLX—————nll_delta_vs_E9:prose 0.00470 nats/token; ci95:prose 95% CI 0.00345 to 0.00591; nll_delta_vs_E9:public_code 0.00119 (95% CI 0.00068 to 0.00171)other or no referenceL104 @ cfd4a00
AnonimousA/REAP-320-GGUF publisher comparison buildREAP-256 duo Q3 (publisher's earlier repo)GGUF61.95download————HumanEval-164 pass@1 89.6%; rare_name_probe 7/10; fabrication_rate 42%other or no referenceL52 @ 4bc31ec
AnonimousA/REAP-320-GGUF measured by AnonimousAmodified weightsREAP-320 multi-domainGGUF—————HumanEval 96.3%other or no referenceL22 @ c806923
antirez/gguf measured by nazeshinjiteQ4GGUF74.84resident————nll 0.2906; nll:long 0.1253; dnll +0.0064 (+0.0026..+0.0095)other or no referenceL111 @ 9374dd4
apetersson/Abliterated publisher comparison buildStock Q5 (apetersson/Qwen3.8-Flash-Next-GGUF)GGUF—————IFEval 5/8; MATH-500 6/8; HumanEval+/MBPP+ 10/12; MMLU-Pro slice 6/8other or no referenceL38 @ 89bb1c9
AtomicChat/GGUF measured by sirfyyn4.27 bpwGGUF94.50download————kld:E 0.079 ± 0.004; kld:W 0.071 ± 0.002; kld:K 0.207 ± 0.006; top1_pct:E/W/K 96.0 / 89.9 / 85.7vs BF16L216 @ 1b64f8e
AtomicChat/GGUF measured by sirfyyn3.84 bpwGGUF84.90download————kld:E 0.199 ± 0.009; kld:W 0.185 ± 0.004; kld:K 0.447 ± 0.012; top1_pct:E/W/K 92.6 / 83.6 / 78.6vs BF16L220 @ 1b64f8e
AtomicChat/GGUF measured by agentionaiAD-4.27bpw Q4_K_MGGUF56.11GPU weights0.22880.58%———vs a Q8_0L63 @ aa0a0c3
bartowski/GGUF measured by sirfyynIQ3_XXSGGUF88.00download————kld:E 0.181 ± 0.008; kld:W 0.106 ± 0.003; kld:K 0.292 ± 0.008; top1_pct:E/W/K 92.8 / 87.6 / 82.4vs BF16L219 @ 1b64f8e
beamster/Sushi-2.6bpw measured by beamsteroMLX oQ5e (repo not named on card)MLX90.16GPU weights0.0625 kld92.40%———vs BF16L50 @ 0658b15
beamster/Sushi-2.6bpw measured by beamsterVontra 4-bit g32 (TensorFold; repo not named on card)EXL3 (sushi pack)81.17GPU weights0.2074 kld85.01%———vs BF16L56 @ 0658b15
beamster/Sushi-2.6bpw measured by beamsteroMLX oQ4e (repo not named on card)MLX74.31GPU weights0.1370 kld88.87%———vs BF16L54 @ 0658b15
beamster/Sushi-2.6bpw publisher comparison buildSushi-4bpw (bf16 n-gram table)EXL3 (sushi pack)68.38GPU weights0.0592 kld92.89%———vs BF16L49 @ 0658b15
beamster/Sushi-2.6bpw publisher comparison buildSushi-3bpw (4-bit n-gram table)EXL3 (sushi pack)52.97GPU weights0.1036 kld90.31%——kld:bf16_ngram_table 0.1006vs BF16L52 @ 0658b15
beamster/Sushi-2.6bpw publisher comparison buildSushi-2bpw (4-bit n-gram table, bf16 KV)EXL3 (sushi pack)37.55GPU weights0.2080 kld85.94%———vs BF16L57 @ 0658b15
beamster/Sushi-2bpw measured by beamsteroMLX oQ5eMLX90.16GPU weights0.062592.40%———vs BF16L50 @ cb6c85e
beamster/Sushi-2bpw measured by beamsterVontra 4-bit g32 (TensorFold)EXL3 (sushi pack)81.17GPU weights0.207485.01%———vs BF16L56 @ cb6c85e
beamster/Sushi-2bpw measured by beamsteroMLX oQ4eMLX74.31GPU weights0.137088.87%———vs BF16L54 @ cb6c85e
beamster/Sushi-2bpw publisher comparison buildSushi-4bpw (bf16 n-gram table)EXL3 (sushi pack)68.38GPU weights0.059292.89%———vs BF16L49 @ cb6c85e
beamster/Sushi-2bpw publisher comparison buildSushi-3bpw (4-bit n-gram table)EXL3 (sushi pack)52.97GPU weights0.103690.31%———vs BF16L52 @ cb6c85e
beamster/Sushi-2bpw publisher comparison buildSushi-2.6bpw (4-bit n-gram table)EXL3 (sushi pack)47.19GPU weights0.135589.08%———vs BF16L53 @ cb6c85e
beamster/Sushi-2bpw publisher comparison buildSushi-4bpw with 4-bit n-gram tableEXL3 (sushi pack)—0.0654————vs BF16L80 @ cb6c85e
beamster/Sushi-2bpw publisher comparison buildSushi-2.6bpw with bf16 n-gram tableEXL3 (sushi pack)—0.1303————vs BF16L81 @ cb6c85e
beamster/Sushi-2bpw publisher comparison buildSushi-3bpw with bf16 n-gram tableEXL3 (sushi pack)—0.1006————vs BF16L81 @ cb6c85e
beamster/Sushi-3bpw publisher comparison buildSushi-4bpw (bf16 n-gram table)EXL3 (sushi pack)68.38GPU weights0.059292.89%———vs BF16L49 @ 76bd8f5
beamster/Sushi-3bpw publisher comparison buildSushi-2.6bpw (4-bit n-gram table)EXL3 (sushi pack)47.19GPU weights0.135589.08%———vs BF16L53 @ 76bd8f5
beamster/Sushi-3bpw publisher comparison buildSushi-2bpw (4-bit n-gram table)EXL3 (sushi pack)37.55GPU weights0.208085.94%———vs BF16L57 @ 76bd8f5
beamster/Sushi-3bpw publisher comparison buildSushi-3bpw with bf16 n-gram tableEXL3 (sushi pack)—0.1006————vs BF16L81 @ 76bd8f5
beamster/Sushi-3bpw publisher comparison buildSushi-4bpw with 4-bit n-gram tableEXL3 (sushi pack)—0.0654————vs BF16L80 @ 76bd8f5
beamster/Sushi-4bpw measured by beamsteroMLX oQ5e (repo not named on card)MLX90.16GPU weights0.0625 kld92.40%———vs BF16L53 @ f25c0c4
beamster/Sushi-4bpw measured by beamsterVontra 4-bit g32 (TensorFold; repo not named on card)EXL3 (sushi pack)81.17GPU weights0.2074 kld85.01%———vs BF16L59 @ f25c0c4
beamster/Sushi-4bpw measured by beamsteroMLX oQ4e (repo not named on card)MLX74.31GPU weights0.1370 kld88.87%———vs BF16L57 @ f25c0c4
beamster/Sushi-4bpw measured by chatnerdieSushi-4bpwEXL3 (sushi pack)68.40GPU weights————kld:yarn 0.006137; top1:yarn 98.2%; kld:native 0.003740; top1:native 98.4%; tool_calling 28 / 28vs BF16L61 @ 5329132
beamster/Sushi-4bpw publisher comparison buildSushi-3bpw (4-bit n-gram table)EXL3 (sushi pack)52.97GPU weights0.1036 kld90.31%——kld:bf16_ngram_table 0.1006vs BF16L55 @ f25c0c4
beamster/Sushi-4bpw publisher comparison buildSushi-2.6bpw (4-bit n-gram table)EXL3 (sushi pack)47.19GPU weights0.1355 kld89.08%———vs BF16L56 @ f25c0c4
beamster/Sushi-4bpw publisher comparison buildSushi-2bpw (4-bit n-gram table, bf16 KV)EXL3 (sushi pack)37.55GPU weights0.2080 kld85.94%———vs BF16L60 @ f25c0c4
beamster/Sushi-4bpw measured by SiliconSpeciesSushi-4bpwEXL3 (sushi pack)——91.0%——kld:wikitext 0.0550; kld:reasoning 0.0180; kld:32K 0.0248vs BF16L33 @ 950fcc1
bitlamas/GGUF publisher comparison buildDN4 (first build, requantized from unsloth Q5_1)GGUF———2.9165 ± 0.0363—kld_vs_UD-Q4_K_XL 0.0282; down_err_vs_bf16 0.00667other or no referenceL128 @ 408a8e6
bitlamas/GGUF publisher comparison buildsame file, mainline b11681 as shipped (int8 coopmat on)GGUF——93.71 %——kld_vs_UD-Q4_K_XL 0.0383other or no referenceL151 @ 408a8e6
chatnerdie/EXL3-Sushi publisher comparison buildearlier 4-bit -sh build (different calibration data)EXL3 (sushi pack)—————MMLU-Pro 1,400 q, YaRN 68.64%; MMLU-Pro 1,400 q, native RoPE 68.79%vs BF16L83 @ 5329132
codelight-ai/NVFP4-PLE-INT3 publisher comparison buildINT6 asym per-row (PLE scheme measured, not shipped)NVFP4 experts + INT3 PLE (compressed-tensors style)—————ple_rel_err 0.0241; ple_cos 0.999710other or no referenceL102 @ 59ecf4e
codelight-ai/NVFP4-PLE-INT3 publisher comparison buildINT4 asym per-row (PLE scheme measured, not shipped)NVFP4 experts + INT3 PLE (compressed-tensors style)—————ple_rel_err 0.0958; ple_cos 0.995406other or no referenceL103 @ 59ecf4e
cyankiwi/AWQ-INT4 measured by primitive-aiAWQ INT4 g32AWQ—————tool_calling 78.0 (n=1)other or no referenceL214 @ e1b0da6
cyankiwi/AWQ-INT4 measured by primitive-aiAWQ INT4 g32AWQ—————tool_calling 78.0 (n=1)other or no referenceL251 @ 07915ee
cyankiwi/AWQ-INT4 measured by VnimanieAIAWQ-INT4AWQ—————GPQA paired 40 correct 26/40; GPQA hard-20 correct 10; median completion tokens 3,391other or no referenceL186 @ 142456d
cygnal/Heretic2-IQ4XS-NGQ4-GGUF publisher comparison buildQ6_K_v7_v7GGUF129.0unknown————HumanEval+ base pass@1 88.4%; HumanEval+ plus pass@1 82.9%other or no referenceL86 @ b7a350f
cygnal/Heretic2-IQ4XS-NGQ4-GGUF publisher comparison buildQ6_KGGUF129.0unknown————HumanEval+ base pass@1 88.4%; HumanEval+ plus pass@1 81.7%other or no referenceL87 @ b7a350f
cygnal/Heretic2-IQ4XS-NGQ4-GGUF publisher comparison buildBase flash-next IQ4_XS NGQ4 (not abliterated)GGUF—————HumanEval+ base pass@1 82.3%; HumanEval+ plus pass@1 78.0%other or no referenceL88 @ b7a350f
cygnal/Heretic2-IQ4XS-NGQ4-GGUF publisher comparison buildUncensored IQ4_XS NGQ4GGUF—————HumanEval+ base pass@1 81.7%; HumanEval+ plus pass@1 78.7%other or no referenceL90 @ b7a350f
cygnal/Heretic2-IQ4XS-NGQ4-GGUF publisher comparison buildHeretic2 IQ4_XS run in Halogen (BYO, no overlay)GGUF—————HumanEval+ base pass@1 65.2%; HumanEval+ plus pass@1 61.0%other or no referenceL91 @ b7a350f
Cyronius/131B-A6B-GGUF publisher comparison buildtable fully deleted (0/16)GGUF61.20download——4.80—tools .975; GSM8K-15 1.000; MMLU-30 .767other or no referenceL51 @ 2b7b5c7
ddalcu/MLX-Serve-8bit publisher comparison buildiQ-MLX-4.7bpwMLX (mlx-serve pack)107.0download—————no quality claimL30 @ f61a712
ddalcu/MLX-Serve-8bit publisher comparison buildmixed-4-8bitMLX (mlx-serve pack)107.0download—————no quality claimL31 @ f61a712
ddalcu/MLX-Serve-8bit publisher comparison buildiQ-MLX-3.3bpwMLX (mlx-serve pack)86.00download—————no quality claimL29 @ f61a712
ddalcu/MLX-Serve-iQ-MLX-3.3bpw publisher comparison buildmixed-4-8bitMLX73.00download—89.1%——prob_mass 0.738vs BF16L55 @ 4fabeb5
ddalcu/MLX-Serve-iQ-MLX-3.3bpw measured by beamsteriQ-MLX 3.3bpwMLX54.33GPU weights0.198786.28%———vs BF16L55 @ cb6c85e
ddalcu/MLX-Serve-iQ-MLX-3.3bpw measured by beamsteriQ-MLX 3.3bpwMLX54.33GPU weights0.1987 kld86.28%———vs BF16L55 @ 0658b15
ddalcu/MLX-Serve-iQ-MLX-3.3bpw measured by beamsteriQ-MLX 3.3bpwMLX54.33GPU weights0.1987 kld86.28%———vs BF16L58 @ f25c0c4
ddalcu/MLX-Serve-iQ-MLX-3.3bpw measured by beamsteriQ-MLX 3.3bpwMLX54.33GPU weights0.198786.28%———vs BF16L55 @ 76bd8f5
ddalcu/MLX-Serve-iQ-MLX-3.3bpw measured by showkieoiQ-MLX 3.3bpwMLX54.33GPU weights0.198786.28%———vs BF16L121 @ 851e86b
ddalcu/MLX-Serve-iQ-MLX-4.7bpw publisher comparison buildmixed-4-8bit (no calibration; ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit)MLX (mlx-serve)75.30resident0.158———top1:all 87.9; top1:agent 89.0; top1:code 88.0; top1:math 87.5; top1:prose 87.0; kld_median 0.0066vs BF16L61 @ a0e89db
ddalcu/MLX-Serve-iQ-MLX-4.7bpw measured by airagrpiQ-MLX-4.7bpwMLX—0.129 kld89.8%———vs BF16L121 @ af5fc98
ddalcu/MLX-Serve-mixed-4-8bit measured by beamstermixed-4-8bitMLX75.30GPU weights0.081891.39%———vs BF16L51 @ cb6c85e
ddalcu/MLX-Serve-mixed-4-8bit measured by beamstermixed-4-8bitMLX75.30GPU weights0.0818 kld91.39%———vs BF16L51 @ 0658b15
ddalcu/MLX-Serve-mixed-4-8bit measured by beamstermixed-4-8bitMLX75.30GPU weights0.0818 kld91.39%———vs BF16L54 @ f25c0c4
ddalcu/MLX-Serve-mixed-4-8bit measured by beamstermixed-4-8bitMLX75.30GPU weights0.081891.39%———vs BF16L51 @ 76bd8f5
ddalcu/MLX-Serve-mixed-4-8bit measured by showkieomixed-4-8bitMLX75.30GPU weights0.081891.39%———vs BF16L117 @ 851e86b
ddalcu/MLX-Serve-mixed-4-8bit measured by xrrandmixed-4-8 on mlx-serveMLX—————Tool-Eval-Bench score 85–90 (typ. 88–89)other or no referenceL48 @ afd39de
DeepSeek V4 Flash 0731 (hosted; different model) measured by VnimanieAIhosted—————τ²-Bench Airline pass^1 0.700; MMLU_RU strict 0.837 ± 0.012; GPQA Diamond strict @32k 0.854 ± 0.025other or no referenceL132 @ 142456d
donught/NVFP4-dense publisher comparison buildsame checkpoint on GH200 (Marlin W4A16 fallback)NVFP4—————GSM8K (n=200) 98.5 %other or no referenceL79 @ 8ce6fc7
EliovpAI/W3A8-Paiton-RDNA4 publisher comparison buildINT3-hq (measured, not released)Paiton (W3A8)——91.65 %——kl_free 0.0490; kl_routing_locked 0.0323vs BF16L118 @ e5cee0b
EliovpAI/W3A8-Paiton-RDNA4 publisher comparison buildINT4 (separate W4A8 repo)Paiton (W3A8)——93.51 %——kl_free 0.0288; kl_routing_locked 0.0135vs BF16L119 @ e5cee0b
EliovpAI/W3A8-Paiton-RDNA4 publisher comparison buildMXFP4 OCP RTN (reference measurement, same trunk)Paiton (W3A8)——91.69 %——kl_free 0.0515; kl_routing_locked 0.0320vs BF16L120 @ e5cee0b
EliovpAI/W3A8-Paiton-RDNA4 measured by EliovpAIINT3Paiton (W3A8)——90.74 %——kl_free 0.0604; kl_locked 0.0413vs BF16L100 @ 5e4a8db
EliovpAI/W4A8-Paiton-RDNA4 publisher comparison buildall-INT4, 8-bit trunk, exact arithmeticPaiton W4A8 (INT4/INT3 hybrid experts)——93.51 %——kl_free 0.0288; kl_locked 0.0135vs BF16L101 @ 5e4a8db
EliovpAI/W4A8-Paiton-RDNA4 publisher comparison buildINT4/INT3 hybrid, 8-bit trunkPaiton W4A8 (INT4/INT3 hybrid experts)——93.49 %——kl_free 0.0293; kl_locked 0.0152vs BF16L102 @ 5e4a8db
EliovpAI/W4A8-Paiton-RDNA4 publisher comparison buildstandard MXFP4 RTN (own measurement)Paiton W4A8 (INT4/INT3 hybrid experts)——91.69 %——kl_free 0.0515; kl_locked 0.0320vs BF16L104 @ 5e4a8db
fabiogreter/lily-q4-xl publisher comparison buildfabiogreter/Qwen3.8-Flash-Next-lily-q4 (4-bit sibling)lily qwen4_exp-affine-v1 (safetensors, engine-specific)—————tool_call_continuation:opencode 56K turn 18 of 30other or no referenceL88 @ 0f94842
fuutott/MLX-t5-imatrix-q3down-ple8 publisher comparison buildAffine Q2 + imatrix, all expert projections (stock-oMLX variant)MLX (oMLX fork, ternary)38.23resident0.47776.4 %4.31 (× 1.31)—median_kld 0.186other or no referenceL108 @ 15ec9e8
fuutott/MLX-t5-imatrix-q3down-ple8 publisher comparison buildT5 + imatrix gate/up, Q2 downMLX (oMLX fork, ternary)33.18resident0.57374.0 %4.67 (× 1.42)—median_kld 0.237other or no referenceL109 @ 15ec9e8
fuutott/MLX-t5-imatrix-q3down-ple8 publisher comparison buildT5 weight-only gate/up (no imatrix), Q2 downMLX (oMLX fork, ternary)33.18resident0.69071.3 %5.27 (× 1.60)—median_kld 0.327; MMLU (5-shot) 237 (79.0 %); GSM8K 85 (85.0 %); TruthfulQA MC1 178 (89.0 %); CMMLU (0-shot) 78 (78.0 %); bench_total 578 (82.6 %)other or no referenceL110 @ 15ec9e8
ggml-org/GGUF measured by agentionaiIQ4_NLGGUF73.24GPU weights0.20081.29%———vs a Q8_0L64 @ aa0a0c3
ggml-org/GGUF measured by ivanfioravantiQ8_0GGUF——98.7%——target_logprob_mae 0.0211vs BF16L128 @ 59a55fb
GGUF Q4 (ds4) — publisher/repo not identified on card measured by Ar3ntizGGUF Q4 (ds4)GGUF177.3download0.061———Total 99/120; GSM8K 18; MATH 16; MMLU-Pro 7; ARC-C 19; BBH 19; Long context 20vs BF16L56 @ ba49b29
gpillon/trellis-a25-ignis publisher comparison buildFP8 non-expert weights with BF16 expertstrellis (EXL3/QTIP family, .ninfer)—————MMLU-Pro proxy 74.7%vs BF16L63 @ f75e492
Gufo (stack; repo not named on card) measured by jcbtcGufo—0.027966 kl96.436%——tail_ppl 2.03502; hard_tools_per100 73; hermes_pass1_per100 95; hermes_pass2_per100 95vs BF16L189 @ f3f020e
HaloBox (stack; repo not named on card) measured by jcbtcHaloBox—0.029182 kl96.143%——tail_ppl 2.03642; hard_tools_per100 63; hermes_pass1_per100 90; hermes_pass2_per100 96vs BF16L188 @ f3f020e
Halogen (stack; repo not named on card) measured by jcbtcHalogen—0.085905 kl92.773%——tail_ppl 2.07968; hard_tools_per100 70; hermes_pass1_per100 93; hermes_pass2_per100 92vs BF16L191 @ f3f020e
Inferact/NVFP4 measured by RedHatAINVFP4NVFP4—————GPQA Diamond 91.4; MMLU-Pro 88.0; AIME 2026 100; LiveCodeBench 94.0; Terminal-Bench 2.1 84.5; SWE-bench Verified 79.8; SWE-bench Pro 60.9; DeepSWE 1.1 58.1; recovery:DeepSWE 1.1 88.0; recovery_avg 97.6vs BF16L78 @ c8f2fb1
inferencerlabs/MLX-Q9 publisher comparison buildQ3.5MLX———168.0—token_accuracy_pct 43.45%; missed_divergence_pct 72.57%other or no referenceL32 @ 627f6d6
inferencerlabs/MLX-Q9 publisher comparison buildQ4.5MLX———1.33593—token_accuracy_pct 91.65%; missed_divergence_pct 17.28%other or no referenceL33 @ 627f6d6
inferencerlabs/MLX-Q9 publisher comparison buildQ5.5MLX———1.23437—token_accuracy_pct 95.05%; missed_divergence_pct 17.28%other or no referenceL34 @ 627f6d6
inferencerlabs/MLX-Q9 publisher comparison buildQ6.5MLX———1.21875—token_accuracy_pct 96.65%; missed_divergence_pct 12.03%other or no referenceL35 @ 627f6d6
inferencerlabs/MLX-Q9 publisher comparison buildQ8.5MLX———1.21875—token_accuracy_pct 97.65%; missed_divergence_pct 9.92%other or no referenceL36 @ 627f6d6
Intel/W4A16-AutoRound measured by primitive-aiINT4 g128AutoRound—————tool_calling 80.5 (n=1)other or no referenceL212 @ e1b0da6
Intel/W4A16-AutoRound measured by primitive-aiINT4 g128AutoRound—————tool_calling 80.5 (n=1)other or no referenceL249 @ 07915ee
intel-ai/MXFP4-BF16Attn-MXFP8MTP-CT-AutoRound publisher comparison buildSibling -BF16Attn-1007 (MTP fully bf16)compressed-tensors (MXFP4/MXFP8)—————GSM8K strict 0.9644; GSM8K flexible 0.9636; MMLU 0.8641; PIQA acc 0.8248; PIQA acc_norm 0.8308; HellaSwag acc 0.6856; HellaSwag acc_norm 0.8726vs BF16L85 @ 7810203
intel-ai/MXFP4-BF16Attn-MXFP8MTP-CT-AutoRound publisher comparison buildThis checkpoint with MTP enabled (num_speculative_tokens 1)compressed-tensors (MXFP4/MXFP8)—————GSM8K limit-100 0.98vs BF16L71 @ 7810203
intel-ai/MXFP4-BF16Attn-MXFP8MTP-PLEMXFP8-CT-AutoRound publisher comparison buildSibling -MTP8-1008 (PLE table bf16; intel-ai/Qwen3.8-Flash-Next-MXFP4-BF16Attn-MXFP8MTP-CT-AutoRound)compressed-tensors (MXFP4/MXFP8)180.1download————GSM8K strict 0.9651; GSM8K flexible 0.9644; MMLU 0.8655; PIQA acc 0.8243; PIQA acc_norm 0.8330; HellaSwag acc 0.6845; HellaSwag acc_norm 0.8708vs BF16L100 @ 9008b69
intel-ai/MXFP4-BF16Attn-PLEMXFP8-CT-AutoRound publisher comparison buildSibling -BF16Attn (PLE table bf16)compressed-tensors (MXFP4 experts + MXFP8 PLE)—————GSM8K (strict / flexible) 0.9644 / 0.9636; MMLU 0.8641; PIQA (acc / acc_norm) 0.8248 / 0.8308; HellaSwag (acc / acc_norm) 0.6856 / 0.8726vs BF16L102 @ b17da51
intel-ai/MXFP4-BF16SharedExperts-CT-AutoRound publisher comparison buildSibling variant (shared_expert = MXFP8)compressed-tensors mixed-precision (MXFP4/MXFP8)—————GSM8K strict 0.9621; GSM8K flexible 0.9621; MMLU 0.8626; PIQA acc 0.8226; PIQA acc_norm 0.8270; HellaSwag acc 0.6832; HellaSwag acc_norm 0.8707vs BF16L93 @ a0ed54e
ISTA-DASLab/GSQ-RCO-GGUF measured by TylogiIQ3_SGGUF83.62download0.13717 kld86.90%———reference not statedL95 @ 3cdfba4
ISTA-DASLab/GSQ-RCO-GGUF measured by sirfyynIQ3_SGGUF83.60download————kld:E 0.069 ± 0.004; kld:W 0.061 ± 0.002; kld:K 0.183 ± 0.005; top1_pct E/W/K 96.4 / 90.6 / 86.8vs BF16L86 @ 6bbaae7
ISTA-DASLab/GSQ-RCO-GGUF measured by sirfyynIQ3_SGGUF83.60download————kld:E 0.069 ± 0.004; kld:W 0.061 ± 0.002; kld:K 0.183 ± 0.005; top1_pct:E/W/K 96.4 / 90.6 / 86.8vs BF16L221 @ 1b64f8e
ISTA-DASLab/GSQ-RCO-GGUF measured by TylogiIQ3_XXSGGUF75.84download0.22134 kld83.00%———reference not statedL94 @ 3cdfba4
ISTA-DASLab/GSQ-RCO-GGUF measured by sirfyynIQ3_XXSGGUF75.80download————kld:E 0.118 ± 0.006; kld:W 0.112 ± 0.003; kld:K 0.264 ± 0.007; top1_pct:E/W/K 95.0 / 87.4 / 83.6vs BF16L226 @ 1b64f8e
ISTA-DASLab/GSQ-RCO-GGUF measured by TylogiIQ2_XSGGUF68.03download0.33374 kld79.06%———reference not statedL93 @ 3cdfba4
ISTA-DASLab/GSQ-RCO-GGUF measured by sirfyynIQ2_XSGGUF68.00download————kld:E 0.179 ± 0.007; kld:W 0.191 ± 0.004; kld:K 0.382 ± 0.010; top1_pct:E/W/K 93.0 / 83.6 / 80.3vs BF16L229 @ 1b64f8e
ISTA-DASLab/GSQ-RCO-GGUF measured by TylogiQ2_0GGUF66.42download0.40660 kld77.27%———reference not statedL92 @ 3cdfba4
ISTA-DASLab/GSQ-RCO-GGUF measured by sirfyynQ2_0GGUF66.40download————kld:E 0.223 ± 0.009; kld:W 0.236 ± 0.005; kld:K 0.485 ± 0.012; top1_pct:E/W/K 92.4 / 82.3 / 77.7vs BF16L230 @ 1b64f8e
ISTA-DASLab/GSQ-RCO-GGUF measured by Maximilian228IQ3_SGGUF55.83RAM in use—90.0%———vs a Q8_0L39 @ 379ff0b
ISTA-DASLab/GSQ-RCO-GGUF measured by Maximilian228IQ3_SGGUF55.83RAM in use—90.0%———other or no referenceL46 @ 9ac833e
ISTA-DASLab/GSQ-RCO-GGUF measured by Maximilian228IQ3_SGGUF55.83RAM in use—90.0%———vs a Q8_0L43 @ d8fc7df
ISTA-DASLab/GSQ-RCO-GGUF measured by lackonendesIQ3_SGGUF54.87trunk only————kl:code 0.102; kl:plain-English 0.085vs BF16L125 @ 257599e
ISTA-DASLab/GSQ-RCO-GGUF measured by Maximilian228IQ3_XXSGGUF48.32RAM in use—87.3%———vs a Q8_0L40 @ 379ff0b
ISTA-DASLab/GSQ-RCO-GGUF measured by Maximilian228IQ3_XXSGGUF48.32RAM in use—87.3%———other or no referenceL47 @ 9ac833e
ISTA-DASLab/GSQ-RCO-GGUF measured by Maximilian228IQ3_XXSGGUF48.32RAM in use—87.3%———vs a Q8_0L44 @ d8fc7df
ISTA-DASLab/GSQ-RCO-GGUF measured by agentionaiIQ3_XXSGGUF47.03GPU weights0.31177.68%———vs a Q8_0L61 @ aa0a0c3
ISTA-DASLab/GSQ-RCO-GGUF measured by lackonendesIQ3_XXSGGUF47.03trunk only————kl:code 0.172; kl:plain-English 0.139vs BF16L126 @ 257599e
ISTA-DASLab/GSQ-RCO-GGUF measured by Maximilian228IQ2_XSGGUF40.80RAM in use—85.0%———vs a Q8_0L41 @ 379ff0b
ISTA-DASLab/GSQ-RCO-GGUF measured by Maximilian228IQ2_XSGGUF40.80RAM in use—85.0%———other or no referenceL48 @ 9ac833e
ISTA-DASLab/GSQ-RCO-GGUF measured by Maximilian228IQ2_XSGGUF40.80RAM in use—85.0%———vs a Q8_0L45 @ d8fc7df
ISTA-DASLab/GSQ-RCO-GGUF measured by Maximilian228Q2_0GGUF39.73RAM in use—84.1%———vs a Q8_0L42 @ 379ff0b
ISTA-DASLab/GSQ-RCO-GGUF measured by Maximilian228Q2_0GGUF39.73RAM in use—84.1%———other or no referenceL49 @ 9ac833e
ISTA-DASLab/GSQ-RCO-GGUF measured by Maximilian228Q2_0GGUF39.73RAM in use—84.1%———vs a Q8_0L46 @ d8fc7df
ISTA-DASLab/GSQ-RCO-GGUF measured by agentionaiIQ2_XSGGUF39.21GPU weights0.42074.22%———vs a Q8_0L60 @ aa0a0c3
ISTA-DASLab/GSQ-RCO-GGUF measured by WarlaxZIQ2_XS (in Strata 0.1.38)GGUF—————ppl:Strata prompts 7.8371; ppl:Infernix prompts 5.0518other or no referenceL130 @ c7dfb96
ISTA-DASLab/GSQ-RCO-GGUF measured by Jakevinrepo not named on cardISTA 2-bit (file not specified)GGUF—————kl_nats_per_token 0.217; mcq_correct_of_216 162other or no referenceL30 @ e5f8a6b
ISTA-DASLab/GSQ-RCO-GGUF quoted by sleepyeldrazi, not measuredIQ3_XXS (base pack)GGUF———3.9044 ppl:wikitext——reference not statedL62 @ cb186e7
ivanfioravanti/DS4-IQ2 publisher comparison buildprevious MXFP4-down buildGGUF——91.33% top1_pct_vs_bf16——target_nll 0.290735; first_token_match 69/99; hard_smoke_12q 6 correct, 0 wrong, 6 incompletevs BF16L106 @ b8b2039
ivanfioravanti/DS4-Q4 publisher comparison buildprevious Q4_0-routed releaseGGUF——96.5%——target_logprob_mae 0.0430vs BF16L129 @ 59a55fb
Jakevin/ternary-mlx publisher comparison buildall-ternary allocation (not released)MLX (mlx-ternary mixed)—————pct_of_FP8_score 57%other or no referenceL38 @ e5f8a6b
jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4 publisher comparison buildCIRU v4.4.1 (retained)GGUF (custom CIRU IU4 format)79.40download0.056117 kl93.750%——tail_ppl 2.05516; hard_tools_per100 67; hermes_pass1_per100 89; hermes_pass2_per100 88vs BF16L186 @ f3f020e
jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4 publisher comparison buildv4 (historical, earlier runtime/settings)GGUF (custom CIRU IU4 format)—————HumanEval 160/164; HumanEval+ 155/164; ARC-Challenge 1,143/1,172; ToolEval Standard 115/138; ToolEval Hard 23/30vs BF16L276 @ f3f020e
jcbtc/Qwen3.8-Flash-CIRU-STRIX-Orca publisher comparison buildOriginal Orca GGUF (retained parts)GGUF (CIRU compact expert format)79.40download——2.03421—nll 0.710105other or no referenceL217 @ 139c9f5
ji-farthing/ik-llama-GGUF publisher comparison buildIQ4_KT_v2 (ji-farthing/..._v2 repo)GGUF86.67download0.132805 kld:english86.276% top1:english3.5335 ppl:english—kld:code 0.124558; top1:code 93.216%; ppl:code 1.8031vs BF16L112 @ 8fb1860
ji-farthing/ik-llama-GGUF publisher comparison buildIQ4_KT (replaced)GGUF—0.1325 kld:english85.74% top1:english3.5625 ppl:english—kld:code 0.1247; top1:code 92.41%; ppl:code 1.8209vs BF16L113 @ 8fb1860
ji-farthing/ik-llama-GGUF publisher comparison buildIQ3_KT (replaced)GGUF—0.1515 kld:english85.05% top1:english3.6119 ppl:english—kld:code 0.1350; top1:code 92.13%; ppl:code 1.8273vs BF16L114 @ 8fb1860
ji-farthing/ik-llama-GGUF_v2 publisher comparison buildPrevious IQ4_KT with IQ4_NL fallbacksGGUF94.94download0.132548 kld:english85.738% top1:english——kld:code 0.124749; top1:code 92.414%vs BF16L72 @ ca863d1
ji-farthing/ik-llama-GGUF_v2 publisher comparison buildPrevious IQ3_KT with IQ4_NL fallbacksGGUF86.14download0.151454 kld:english85.054% top1:english——kld:code 0.134990; top1:code 92.131%vs BF16L71 @ ca863d1
jomcgi-org/NVFP4-oominf publisher comparison buildsame files, max-perf mode (--dense fp8 --expert-precision bf16)NVFP4—————GSM8K (300, t0) 97.00; MMLU-Pro (200) 77.50; RULER 8k+32k 99.62; RULER 95k 98.08vs BF16L113 @ 84010e0
JRamirez-UAB/EXL3-2.25bpw-48GB publisher comparison buildsame recipe with exllamav3 default calibrationEXL3—————kld public/internal 0.051 / 0.058vs BF16L127 @ 9eb541a
Jundot/oQ4e-mtp measured by umixeruniform oQ4eMLX106.3download—————other or no referenceL62 @ 73d2cd0
Jundot/oQ4e-mtp publisher comparison buildoQ4e (previous, 2026-08-27)MLX—————kl:code 0.0877; kl:novel 0.1909vs BF16L34 @ c4ddfb5
Jundot/oQ4e-mtp quoted by mensaprodigy, not measuredoQ4e MTP (oMLX records, quoted)MLX—————HumanEval 152/164 = 92.7%; LiveCodeBench 55/100other or no referenceL72 @ 0ee18f7
Jundot/oQ4e-mtp quoted by mensaprodigy, not measuredoQ4e MTP (oMLX records, quoted)MLX—————HumanEval 152/164 = 92.7%; LiveCodeBench (Thinking Yes, 100) 55.0%other or no referenceL47 @ ff32837
Jundot/oQ4e-mtp measured by xrrandoQ4e on oMLXMLX—————Tool-Eval-Bench score 92–93other or no referenceL46 @ afd39de
Jundot/oQ4e-mtp measured by GBP-DEoQ4eMLX—————test:8-person constraint puzzle (thinking off) Failed; test:five-trigger state machine (thinking on) did not complete the required answer; test:Python debugging (thinking on) never produced the required final answer; test:four-tank state machine Exact; test:Boolean/SAT Exactother or no referenceL144 @ 387311f
Jundot/oQ6e-mtp quoted by umixer, not measuredoQ6eMLX150.5download—————other or no referenceL86 @ 73d2cd0
Kashapps/Uncensored-ROCmFP4-FAST-GGUF publisher comparison buildaligned build (same tier)GGUF (ROCmFP4, fork-only)—————refusal:Harmful (24) 0 comply; Harmless (12) 10 ok; quality:Quality (8) 6/8other or no referenceL148 @ 1aa1c09
kingjones777/ROCmFP4-STRIX-GGUF publisher comparison buildSTRIX_LEAN tierGGUF (ROCmFP4, Q4_0_ROCMFP4_STRIX)105.8download—————no quality claimL205 @ 7320914
kingjones777/ROCmFP4-STRIX-GGUF publisher comparison buildFAST tierGGUF (ROCmFP4, Q4_0_ROCMFP4_STRIX)94.42download—————no quality claimL204 @ 7320914
kingjones777/ROCmFP4-STRIX_LEAN-GGUF publisher comparison buildSTRIX tierGGUF (ROCmFP4)121.8download—————no quality claimL206 @ bf7fbc6
kingjones777/ROCmFP4-STRIX_LEAN-GGUF publisher comparison buildFAST tierGGUF (ROCmFP4)94.42download—————no quality claimL204 @ bf7fbc6
kingjones777/Uncensored-ROCmFP4-FAST-GGUF measured by kingjones777modified weightsplain FAST (no imatrix)GGUF94.38download——5.3465 ± 0.034——other or no referenceL57 @ 9abf014
kingjones777/Uncensored-ROCmFP4-FAST-GGUF publisher comparison buildaligned build, same tier (publisher's own)GGUF (ROCmFP4)—————quality_set 6/8; harmless_ok 10; harmful_comply 0other or no referenceL213 @ ce1bddb
kingjones777/Uncensored-ROCmFP4-STRIX_LEAN-GGUF publisher comparison buildaligned build (same tier)GGUF (ROCmFP4, fork-only)—————refusal:Harmful (24) 0 comply; Harmless (12) 10 ok; quality:Quality (8) 6/8other or no referenceL210 @ a4d9d7b
kingjones777/Uncensored-ROCmFP4-STRIX_LEAN-imatrix-GGUF publisher comparison buildplain FAST (no imatrix), separate repoGGUF———5.3465 ± 0.034——other or no referenceL56 @ 503e614
kingjones777/Uncensored-ROCmFP4-STRIX_LEAN-imatrix-GGUF publisher comparison buildFAST imatrix, separate repoGGUF———5.0337 ± 0.031——other or no referenceL57 @ 503e614
leoncca/AWQ-g32 publisher comparison buildprevious release of this repoAWQ——96.32%——kld_top20_tail 0.0305; kld_bounds [0.030235, 0.327424]vs BF16L136 @ ff2fb53
Litwein/REAP320-oQ3e-DWQ-MTP-Vision-MLX publisher comparison buildDWQ v7MLX—————ce_gap:Overall 0.0934; kl20:Overall 0.1065vs BF16L139 @ 9a16783
Litwein/REAP320-oQ3e-DWQ-MTP-Vision-MLX publisher comparison buildDWQ v8MLX—————ce_gap:Overall 0.0921; kl20:Overall 0.1052; ce_gap:independent_gate 0.0986; kl20:independent_gate 0.0845; ce_gap:vision 0.2637; kl20:vision 0.2342vs BF16L139 @ 9a16783
Litwein/REAP320-oQ3e-fp16-DWQ-MTP-Vision-MLX publisher comparison buildBF16-typed sibling (Litwein/Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX)MLX (oMLX)70.80download—————vs BF16L212 @ 46a83dc
Litwein/REAP320-oQ3e-fp16-DWQ-MTP-Vision-MLX publisher comparison buildv7 (earlier DWQ round)MLX (oMLX)—————ce_gap:Overall 0.0934; kl20:Overall 0.1065vs BF16L194 @ 46a83dc
Litwein/REAP320-oQ3e-fp16-DWQ-MTP-Vision-MLX publisher comparison buildv8 (earlier DWQ round)MLX (oMLX)—————ce_gap:Overall 0.0921; kl20:Overall 0.1052vs BF16L194 @ 46a83dc
local-inference-lab/NVFP4 measured by octolixmodified weightsNVFP4 @ 7c4f1bc1 on TensorFold 0.6.0NVFP4—————GSM8K (250) 246; HumanEval pass@1 (164) 158other or no referenceL105 @ a713d8c
LOKSUN-158/SamQuant-Uncensored-MLX-Serve-Mix-C58 publisher comparison buildAp73 (same-budget all-4-bit pack)MLX (mlx-serve mixed)——85.76%——top1_delta_pts +0.36vs BF16L87 @ a6692a8
Lottolabs/Coder-P150 publisher comparison buildIQ3_S residents, min-MSE BFP4 of the GGUF values (no GPTQ)BFP4 (Tenstorrent)—.0437 kl——-2.21%top1_agreement .925other or no referenceL73 @ 69e738b
lvkaokao/W4A16-RTN-AutoRound measured by primitive-aiINT4 g128 RTNAutoRound—————tool_calling 80.0 (n=1)other or no referenceL213 @ e1b0da6
lvkaokao/W4A16-RTN-AutoRound measured by primitive-aiINT4 g128 RTNAutoRound—————tool_calling 80.0 (n=1)other or no referenceL250 @ 07915ee
manjunathshiva/tq4a-tq2e-g64 publisher comparison buildsame build before 2026-09-14 rebuildMLX (TurboQuant)———9.92—battery:8_tasks_thinking_disabled 8/8 tasks pass; battery:thinking_enabled_close_within_5000 13/15other or no referenceL224 @ 34811dc
matteiuspi/W4A16-G128-300i publisher comparison buildopt-in dynamic-W8A8 PLE projection (runtime option, one-token shape)ModelSlim INT4 W4 g128 (Ascend 310P)—————rel_l2_err 0.58%; cos 0.999983other or no referenceL152 @ d32978c
matteiuspi/W8A8-DYNAMIC-300i quoted by matteiuspi, not measuredCompanion W4 G128 checkpoint (matteiuspi/Qwen3.8-Flash-Next-W4A16-G128-300i)Ascend ModelSlim W8A8_DYNAMIC (ascend_v1)—————GPQA Diamond (198) 140/198 (70.71%); GPQA Diamond contains correct answer 114/198vs BF16L97 @ 27e5cdc
Maximilian228/NVFP4-Q8-Mix-Strata publisher comparison buildQ8_0 Strata (reference; Maximilian228/Qwen3.8-Flash-Next-Q8_0-Strata)NVFP4+Q8_0 (Strata pack)113.8RAM in use—100%———other or no referenceL40 @ 9ac833e
Maximilian228/NVFP4-Q8-Mix-Strata measured by Maximilian228NVFP4 + Q8_0 mixNVFP4112.7RAM in use—95.0%———vs a Q8_0L34 @ 379ff0b
Maximilian228/NVFP4-Q8-Mix-Strata publisher comparison buildNVFP4 GPTQ Strata (Maximilian228/Qwen3.8-Flash-Next-NVFP4-GPTQ-Strata)NVFP4+Q8_0 (Strata pack)74.09RAM in use0.0379 mean_kl93.6%——top1:thinking 96.5%; kl:thinking 0.0124; ppl_vs_ref 100.1%other or no referenceL43 @ 9ac833e
Maximilian228/Q8_0-Strata measured by Maximilian228Q8_0 (reference)Strata pack (Q8_0 experts)113.8RAM in use—100%———vs a Q8_0L33 @ 379ff0b
Maximilian228/Q8_0-Strata publisher comparison buildNVFP4 + Q8_0 mix (qwen-nvfp4-mix)Strata pack (Q8_0 experts)112.7RAM in use—95.0%———vs a Q8_0L38 @ d8fc7df
Maximilian228/Q8_0-Strata publisher comparison buildNVFP4 GPTQ (qwen-nvfp4-gptq)Strata pack (Q8_0 experts)74.09RAM in use—93.6%———vs a Q8_0L40 @ d8fc7df
Maximilian228/Q8_0-Strata publisher comparison buildNVIDIA NVFP4, converted by us (qwen-nvidia-nvfp4)Strata pack (Q8_0 experts)74.09RAM in use—92.9%———vs a Q8_0L41 @ d8fc7df
mbehr90/fp8 publisher comparison buildNVFP4 (mbehr90/Qwen3.8-Flash-Next-nvfp4)FP8 (compressed-tensors, FP8_DYNAMIC)182.8download————GSM8K (200 q, chat) 0.960; mean_answer_chars 1087vs BF16L30 @ e70592b
mbehr90/nvfp4 publisher comparison buildFP8 (mbehr90/Qwen3.8-Flash-Next-fp8)NVFP4A16 (compressed-tensors)237.1download————GSM8K (200 q, chat) 0.965; mean_answer_chars 1146vs BF16L29 @ d28eef6
meldlabs/Meld-Turbo-Quality-GGUF publisher comparison buildMeld Turbo fastGGUF (EXL3 expert tensors)———2.826—GSM8K (200) 98.0%; HumanEval (164) 93.3%; MMLU (200) 82.5%vs a Q8_0L44 @ ecbcece
MLX affine 4-bit — publisher/repo not identified on card measured by Ar3ntizMLX affine 4-bitMLX—0.202————vs BF16L57 @ ba49b29
myllmbox/hibrid47 publisher comparison buildhibrid46 (int3 PLE table)NVFP4 mixed (vLLM)—————render_gauntlet about half/halfother or no referenceL31 @ 7b83fa0
nazeshinjite/Q8-UDQ4K-DS4 publisher comparison buildbase antirez Q4 (this box)GGUF (ds4)—————nll 0.2906; api_top1 92.08%; topn_recall 88.62%; pair_order_agree 91.58%; top5_mae 0.0985; greedy_first_token 84/100other or no referenceL93 @ 9374dd4
nitinpanj/qwen38-flash-next-v3 publisher comparison buildunspliced checkpointGGUF———5.2777—draft_acceptance 0.751other or no referenceL163 @ e298205
nota-ai/Nota-NVFP4 measured by primitive-aiNVFP4 calibratedNVFP4—————tool_calling 78.0 (n=1); knowledge 92.1; knowledge_completed 92.6; finished_pct 99.3%; tool_call 83.8; tool_abstain 55.0other or no referenceL216 @ e1b0da6
nota-ai/Nota-NVFP4 measured by primitive-aiNVFP4 calibratedNVFP4—————tool_calling 78.0 (n=1); knowledge 92.1; knowledge_completed 92.6; finished_pct 99.3%; tool_call 83.8; tool_abstain 55.0other or no referenceL253 @ 07915ee
NVFP4 (checkpoint not named) measured by turboderpNVFP4 W4A16NVFP4—0.0100—1.3970——vs BF16L27 @ c1def42
NVFP4 (checkpoint not named) measured by turboderpNVFP4 W4A4NVFP4—0.0241—1.4199——vs BF16L27 @ c1def42
nvidia/NVFP4 quoted by intel-ai, not measuredNVFP4compressed-tensors (MXFP4/MXFP8)132.7download—————vs BF16L27 @ 9008b69
nvidia/NVFP4 quoted by intel-ai, not measuredNVFP4compressed-tensors (MXFP4 experts + MXFP8 PLE)132.7download—————vs BF16L26 @ b17da51
nvidia/NVFP4 quoted by intel-ai, not measuredNVFP4 (ModelOpt)compressed-tensors mixed-precision (MXFP4/MXFP8)132.7download—————vs BF16L113 @ a0ed54e
nvidia/NVFP4 quoted by intel-ai, not measuredNVFP4compressed-tensors (MXFP4)132.7download—————vs BF16L106 @ 2855a00
nvidia/NVFP4 measured by Maximilian228NVFP4 (converted to Strata)NVFP474.09RAM in use0.03892.9%——expert_rel_err_vs_bf16 0.01252vs a Q8_0L37 @ 379ff0b
nvidia/NVFP4 measured by Maximilian228NVFP4 (converted to Strata by publisher)NVFP474.09RAM in use—92.9%———other or no referenceL44 @ 9ac833e
nvidia/NVFP4 measured by local-inference-labNVFP4NVFP4—————AA-LCR 74.1; GPQA Diamond 91.5; IFBench 81.0; Tool Eval —other or no referenceL109 @ 6c01ac3
nvidia/NVFP4 measured by primitive-aiNVFP4 (experts only, with primitive's INT4 PLE table, MTP not loaded)NVFP4—————tool_calling 78.3 (n=2); knowledge 92.1other or no referenceL217 @ e1b0da6
nvidia/NVFP4 measured by primitive-aiNVFP4 (experts only, with primitive's INT4 PLE table, MTP not loaded)NVFP4—————tool_calling 78.3 (n=2); knowledge 92.1other or no referenceL254 @ 07915ee
nvidia/NVFP4 measured by SlominNVFP4 (512 experts, run by publisher on DGX Spark under vLLM)NVFP4—————dev:LiveCodeBench (144) 118; dev:16 programming languages (640 tasks) 621; dev:8 human languages (160) 159; dev:tool-call suite (18) 18; dev:HumanEval+ / MBPP+ (264) 238; dev:BFCL v3 (200) 164other or no referenceL118 @ a96007b
nvidia/NVFP4 measured by RedHatAINVFP4NVFP4—————GPQA Diamond 92.4; MMLU-Pro 88.0; AIME 2026 100; LiveCodeBench 95.0; Terminal-Bench 2.1 86.4; SWE-bench Verified 79.8; SWE-bench Pro 62.8; DeepSWE 1.1 56.7; recovery:DeepSWE 1.1 85.9; recovery_avg 98.0vs BF16L79 @ c8f2fb1
omlx 4-bit build of the base model (repo not identified on card) quoted by zankich, not measuredomlx 4-bitcompressed-tensors W4A16 (GPTQ) + INT4 PLE + FP8 KV scales—————MMLU-Pro (300 q subset) 0.843other or no referenceL99 @ ba8fff3
oMLX OptSpeed 4-bit affine 'port' (repo not identified on card) measured by SiliconSpeciesport (oMLX OptSpeed, 4-bit affine)MLX——86.1%——kld:wikitext 0.1416; kld:reasoning 0.0528; kld:32K 0.0841vs BF16L32 @ 950fcc1
oMLX oQ4e (checkpoint not named) measured by beamsteroQ4eMLX74.31GPU weights0.137088.87%———vs BF16L54 @ 76bd8f5
oMLX oQ4e (checkpoint not named) measured by showkieooQ4eMLX74.31GPU weights0.137088.87%———vs BF16L119 @ 851e86b
oMLX oQ5e (checkpoint not named) measured by beamsteroQ5eMLX90.16GPU weights0.062592.40%———vs BF16L50 @ 76bd8f5
oMLX oQ5e (checkpoint not named) measured by showkieooQ5eMLX90.16GPU weights0.062592.40%———vs BF16L115 @ 851e86b
OsaurusAI/Allosaurus-v0.1-125B-A6B-JANGH2 publisher comparison buildJANG_1L (previous release)MLX (JANGH)43.81resident————median_kl:general 0.1430; mean_kl:general 0.838; top1:general 72.09%; median_kl:reasoning 0.0243; top1:reasoning 88.16%; median_kl:tools 0.0050; top1:tools 80.25%; median_kl:office 0.0114; top1:office 86.63%vs BF16L73 @ 01d629e
pentacoxian-dev/IQ3E-Q8D-MTP-GGUF publisher comparison buildsame file, v2 patch tensor split (kernel race)GGUF———2.8608——other or no referenceL316 @ 75e6411
peonist-ai/halogen- measured by Quat3rnionstock v2halogen .hgn (custom)———3.144——other or no referenceL371 @ 2fd9605
peonist-ai/halogen- measured by cygnalHalogen Official w4b (native + overlay)halogen .hgn (custom)—————HumanEval+ base pass@1 82.3%; HumanEval+ plus pass@1 78.0%other or no referenceL89 @ b7a350f
pipenetwork/MLX-4bit publisher comparison build8bitMLX192.2unknown——4.4749—dnll_vs_bf16 +0.0009 [−0.0003, +0.0021]; windows_worse 73/145vs BF16L80 @ aa7c790
pipenetwork/MLX-4bit publisher comparison build6bit (pipenetwork/Qwen3.8-Flash-Next-MLX-6bit)MLX148.0unknown——4.4767—dnll_vs_bf16 +0.0013 [−0.0003, +0.0029]; windows_worse 81/145vs BF16L81 @ aa7c790
pipenetwork/MLX-4bit publisher comparison buildmixed-4_8bitMLX106.2unknown——4.5286+1.3%dnll_vs_bf16 +0.0128 [+0.0109, +0.0148]; windows_worse 128/145vs BF16L82 @ aa7c790
pipenetwork/MLX-4bit publisher comparison build4-bit + hyper-connection read gates at 8-bit (unpublished ablation)MLX———4.9744+11.3%—vs BF16L98 @ aa7c790
pipenetwork/MLX-4bit publisher comparison build4-bit + attention/DeltaNet/shared experts/PLE projections at 8-bit (unpublished ablation)MLX———4.8969+9.5%—vs BF16L99 @ aa7c790
pipenetwork/MLX-4bit publisher comparison build4-bit + embed_tokens and lm_head at 8-bit (unpublished ablation)MLX———5.2843+18.2%—vs BF16L100 @ aa7c790
pipenetwork/MLX-6bit publisher comparison build8bitMLX192.2download——4.4749—dnll_vs_bf16 +0.0009 [−0.0003, +0.0021]; windows_worse 73/145vs BF16L80 @ b4ccbdb
pipenetwork/MLX-6bit publisher comparison buildmixed-4_8bitMLX106.2download——4.5286—dnll_vs_bf16 +0.0128 [+0.0109, +0.0148]; windows_worse 128/145vs BF16L82 @ b4ccbdb
pipenetwork/MLX-6bit publisher comparison build4bitMLX103.8download——5.3914—dnll_vs_bf16 +0.1872 [+0.1778, +0.1968]; windows_worse 145/145vs BF16L83 @ b4ccbdb
pipenetwork/MLX-8bit publisher comparison build6bit (pipenetwork/Qwen3.8-Flash-Next-MLX-6bit)MLX148.0unknown——4.4767—dnll_vs_bf16 +0.0013 [−0.0003, +0.0029]; windows_worse 81/145vs BF16L81 @ f1c23d4
pipenetwork/MLX-8bit publisher comparison buildmixed-4_8bitMLX106.2unknown——4.5286+1.3%dnll_vs_bf16 +0.0128 [+0.0109, +0.0148]; windows_worse 128/145vs BF16L82 @ f1c23d4
pipenetwork/MLX-8bit publisher comparison build4bitMLX103.8unknown——5.3914+20.6%dnll_vs_bf16 +0.1872 [+0.1778, +0.1968]; windows_worse 145/145vs BF16L83 @ f1c23d4
pipenetwork/MLX-8bit publisher comparison build4-bit + hyper-connection read gates at 8-bit (unpublished ablation)MLX———4.9744+11.3%—vs BF16L98 @ f1c23d4
pipenetwork/MLX-8bit publisher comparison build4-bit + attention/DeltaNet/shared experts/PLE projections at 8-bit (unpublished ablation)MLX———4.8969+9.5%—vs BF16L99 @ f1c23d4
pipenetwork/MLX-8bit publisher comparison build4-bit + embed_tokens and lm_head at 8-bit (unpublished ablation)MLX———5.2843+18.2%—vs BF16L100 @ f1c23d4
pipenetwork/MLX-mixed-4_8bit publisher comparison build8bitMLX192.2unknown——4.4749—dnll_vs_bf16 +0.0009 [−0.0003, +0.0021]; windows_worse 73/145vs BF16L80 @ b2c422f
pipenetwork/MLX-mixed-4_8bit publisher comparison build6bit (pipenetwork/Qwen3.8-Flash-Next-MLX-6bit)MLX148.0unknown——4.4767—dnll_vs_bf16 +0.0013 [−0.0003, +0.0029]; windows_worse 81/145vs BF16L81 @ b2c422f
pipenetwork/MLX-mixed-4_8bit publisher comparison build4bitMLX103.8unknown——5.3914+20.6%dnll_vs_bf16 +0.1872 [+0.1778, +0.1968]; windows_worse 145/145vs BF16L83 @ b2c422f
pipenetwork/MLX-mixed-4_8bit publisher comparison build4-bit + hyper-connection read gates at 8-bit (unpublished ablation)MLX———4.9744+11.3%—vs BF16L98 @ b2c422f
pipenetwork/MLX-mixed-4_8bit publisher comparison build4-bit + attention/DeltaNet/shared experts/PLE projections at 8-bit (unpublished ablation)MLX———4.8969+9.5%—vs BF16L99 @ b2c422f
pipenetwork/MLX-mixed-4_8bit publisher comparison build4-bit + embed_tokens and lm_head at 8-bit (unpublished ablation)MLX———5.2843+18.2%—vs BF16L100 @ b2c422f
primitive-ai/NVFP4 publisher comparison buildmixed-NVFP4-FP8 siblingNVFP4183.7unknown————overall 90.3; knowledge 92.2; tool_call 84.8; tool_abstain 56.7; tool_calling 79.2 (n=3)other or no referenceL349 @ e1b0da6
primitive-ai/NVFP4 publisher comparison buildPLE table FP8 per-row / INT4 g16 / NVFP4 g16 sidecars (table sizes 49 / 32 / 28.8 GB)NVFP4—————knowledge 92.2 / 92.9 / 92.2; tool_calling 77.7 / 78.2 / 78.7other or no referenceL163 @ e1b0da6
primitive-ai/mixed-NVFP4-FP8 publisher comparison buildprimitive-ai plain NVFP4NVFP4+FP8186.4unknown————overall 90.2; knowledge 92.2; tool_call 84.6; tool_abstain 56.7; finished_pct 99.4%other or no referenceL324 @ 07915ee
primitive-ai/mixed-NVFP4-FP8 publisher comparison buildv1 of this repo (QSA-only FP8)NVFP4+FP8185.8unknown————overall 90.3; knowledge 92.2; tool_call 85.0; tool_abstain 56.7; finished_pct 99.5%other or no referenceL323 @ 07915ee
primitive-ai/mixed-NVFP4-FP8 measured by donughtmixed NVFP4/FP8NVFP482.70resident————GSM8K (n=200) 97.5 %; mean_completion_tokens_gsm8k 378.9other or no referenceL46 @ 8ce6fc7
primitive-ai/mixed-NVFP4-FP8 publisher comparison buildMTP draft experts FP8 overlayNVFP4+FP8—————tool_calling 78.0other or no referenceL225 @ 07915ee
primitive-ai/mixed-NVFP4-FP8 publisher comparison buildMTP draft experts NVFP4 overlayNVFP4+FP8—————tool_calling 79.0other or no referenceL226 @ 07915ee
pugant/ROCMFP4_STRIX_LEAN-GGUF publisher comparison buildFP2MIX v1 (superseded, FP2 PLE)GGUF57.04download————ppl:Italian holdout (Dante) 37.9027; ppl:EN calibration 7.2928other or no referenceL320 @ a41ebff
pugant/ROCMFP4_STRIX_LEAN-GGUF publisher comparison buildprevious LEAN build (FP8->Q8_0->LEAN requant path)GGUF—————ppl:Italian holdout (Dante) 1.1843; ppl:EN calibration 3.3601other or no referenceL125 @ a41ebff
Quat3rnion/halogen-v2-abliterated publisher comparison buildr2 (historical, revision e7783c58)HGN (Halogen)88.74download——3.134—kl_vs_stock_v2 0.060; top1_vs_stock_v2 92.2%other or no referenceL373 @ 2fd9605
Quat3rnion/halogen-v2-abliterated publisher comparison buildr1 (historical)HGN (Halogen)———3.219—kl_vs_stock_v2 0.129; top1_vs_stock_v2 88.4%other or no referenceL372 @ 2fd9605
Qwen3.8-27B INT8 W8A16 (publisher's previous production model) measured by aixiaomaINT8 W8A16INT8—————total 48/50; arith 20/20; json 8/10; nth 10/10; sort 10/10other or no referenceL96 @ 75234a1
RadixArk/NVFP4 measured by codelight-aiFP8 e4m3 per-tensor PLENVFP4135.2download————ple_rel_err 0.0265; ple_cos 0.999654other or no referenceL101 @ 59ecf4e
RadixArk/NVFP4 quoted by jomcgi-org, not measuredNVFP4 on SGLang (RadixArk-reported, quoted)NVFP4—————GSM8K 97.27; AIME26 98.75vs BF16L90 @ 84010e0
Sawfwair/MLX-4bit measured by sh0wieBase Q4 (512 experts)MLX98.00download————HumanEval pass@1 93.9%other or no referenceL20 @ e3e7da4
Sawfwair/MLX-4bit measured by sh0wieBase Q4 / stock conversion (512 experts)MLX98.00download————HumanEval pass@1 93.9%other or no referenceL20 @ c9bb01b
Sawfwair/MLX-Activation-3bit publisher comparison buildpublished Q4 checkpointMLX——————other or no referenceL72 @ c699bd6
Sawfwair/MLX-Activation-3bit-Native-PLE publisher comparison buildpublished Q4 checkpointMLX—————holdout_peak_memory_bytes 77,180,432,496; mem_reduction_pct 19.58%other or no referenceL82 @ 1cee930
sh0wie/REAP-288-GGUF publisher comparison buildREAP-288 MLX 4-bit (separate repo, same kept experts)GGUF—————HumanEval 91.5%; HumanEval correct 149 of 164other or no referenceL25 @ 2af8d94
sh0wie/REAP-288-MLX-4bit publisher comparison buildREAP-384MLX80.00download————HumanEval pass@1 92.1%other or no referenceL113 @ 668f31b
sh0wie/REAP-288-MLX-4bit publisher comparison buildREAP-320MLX72.00download————HumanEval pass@1 90.9%other or no referenceL114 @ 668f31b
sh0wie/REAP-288-MLX-4bit publisher comparison buildREAP-256MLX65.00download————HumanEval pass@1 88.4%; rare_name_sampling 1 of 10other or no referenceL116 @ 668f31b
sh0wie/REAP-288-MLX-4bit quoted by mensaprodigy, not measuredmodified weightsREAP-288 MLX 4-bit (publisher-reported, quoted)MLX—————HumanEval 91.5%other or no referenceL71 @ 0ee18f7
sh0wie/REAP-288-MLX-4bit publisher comparison buildREAP-288 with top-6 routingMLX—————HumanEval pass@1 84.8%other or no referenceL121 @ 668f31b
sh0wie/REAP-288-MLX-4bit publisher comparison buildREAP-288 with top-4 routingMLX—————HumanEval pass@1 63.4%other or no referenceL121 @ 668f31b
sh0wie/REAP-288-Q8E-MLX publisher comparison buildREAP-288 MLX 4-bit (sh0wie/Qwen3.8-Flash-Next-REAP-288-MLX-4bit)MLX—————HumanEval pass@1 91.5%other or no referenceL45 @ e3e7da4
sh0wie/REAP-384-MLX-4bit publisher comparison buildREAP-320MLX72.00download————HumanEval pass@1 90.9%other or no referenceL94 @ c9bb01b
sh0wie/REAP-384-MLX-4bit publisher comparison buildREAP-288 (sh0wie/Qwen3.8-Flash-Next-REAP-288-MLX-4bit)MLX68.00download————HumanEval pass@1 91.5%other or no referenceL95 @ c9bb01b
sh0wie/REAP-384-MLX-4bit publisher comparison buildREAP-256MLX65.00download————HumanEval pass@1 88.4%other or no referenceL96 @ c9bb01b
sh0wie/REAP-384-MLX-4bit quoted by mensaprodigy, not measuredmodified weightsREAP-384 Q4MLX (oMLX oQ4e)—————HumanEval pass@1 92.1%no quality claimL132 @ ae3a504
sh0wie/REAP-384-MLX-4bit quoted by mensaprodigy, not measuredmodified weightsREAP-384 MLX 4-bit (publisher-reported, quoted)MLX—————HumanEval (164) 92.1%other or no referenceL46 @ ff32837
sh0wie (REAP lineage) quoted by mensaprodigy, not measuredREAP-288 Q4MLX (oMLX oQ4e)—————HumanEval pass@1 91.5%no quality claimL133 @ ae3a504
sh0wie (REAP lineage) quoted by mensaprodigy, not measuredREAP-256 Q4MLX (oMLX oQ4e)—————HumanEval pass@1 88.4%no quality claimL134 @ ae3a504
sh0wie (REAP lineage; card [2] links sh0wie/REAP-384-MLX-4bit) quoted by mensaprodigy, not measuredFull Flash-Next Q4 (512 experts)MLX (oMLX oQ4e)—————HumanEval pass@1 93.9%no quality claimL131 @ ae3a504
showkieo/Sushi-3bpw publisher comparison buildSushi-4bpw (bf16 n-gram table)EXL3 (sushi pack)68.38GPU weights0.063292.99%———vs BF16L116 @ 851e86b
showkieo/Sushi-3bpw publisher comparison buildMLX affine q3 (built for this comparison)EXL3 (sushi pack)58.99GPU weights0.144488.05%———vs BF16L120 @ 851e86b
showkieo/Sushi-3bpw publisher comparison buildSushi-3bpw with bf16 n-gram tableEXL3 (sushi pack)—0.1012————vs BF16L139 @ 851e86b
showkieo/Sushi-3bpw publisher comparison buildSushi-4bpw with 4-bit n-gram tableEXL3 (sushi pack)—0.0666————vs BF16L138 @ 851e86b
SiliconSpecies/EXL3 measured by chatnerdie5bpw-hq (rev ccd64f0)EXL384.07GPU weights————kld:yarn 0.004729; top1:yarn 98.3%; kld:native 0.002070; top1:native 98.7%vs BF16L64 @ 5329132
sirfyyn/FLE-GGUF measured by sirfyynFLE-IQ3_MGGUF88.87download————kld:E 0.051 ± 0.003; kld:W 0.050 ± 0.001; kld:K 0.129 ± 0.004; top1_pct E/W/K 96.9 / 91.5 / 89.1; MKQA German 23.2 %; MKQA English 33.0 %vs BF16L85 @ 6bbaae7
sirfyyn/FLE-GGUF measured by sirfyynFLE-IQ3_XXSGGUF79.48download————kld:E 0.090 ± 0.005; kld:W 0.093 ± 0.002; kld:K 0.227 ± 0.007; top1_pct E/W/K 95.7 / 88.1 / 84.2; MKQA German 21.5 %; MKQA English 30.8 %vs BF16L89 @ 6bbaae7
sirfyyn/FLE-REAP448-GGUF measured by sirfyynmodified weightsFLE-IQ3_M-REAP448GGUF82.00download————kld:E 0.080 ± 0.005; kld:W 0.086 ± 0.003; kld:K 0.185 ± 0.006; top1_pct:E/W/K 96.0 / 89.7 / 87.2vs BF16L222 @ 1b64f8e
sirfyyn/FLE-REAP448-GGUF publisher comparison buildFLE-IQ3_M REAP to 384 experts (not released)GGUF75.13download————MKQA German 17.5 %; MKQA English 21.2 %vs BF16L110 @ 6bbaae7
sirfyyn/FLE-REAP448-GGUF publisher comparison buildFLE-IQ3_M 384 experts by mean activation norm (not released)GGUF75.13download————MKQA German 11.2 %; MKQA English 15.5 %vs BF16L111 @ 6bbaae7
StillDeadcode/qwen3.8-next-flash-fp8-iq4r-moe measured by Dyluhnbaseline container (fp8 / iq4 experts)FP8—0.0537 mean_kl93.6%4.0983 (×1.0051)——vs BF16L42 @ 473882c
SuperPUPS/INT8x17-AttnINT4 publisher comparison buildINT8x25-AttnBF16 (SuperPUPS sibling)compressed-tensors (INT8/INT4 pack-quantized)214.1download—————other or no referenceL33 @ 8d7e6f7
SuperPUPS/INT8x25-AttnBF16 publisher comparison buildINT8x17-AttnINT4compressed-tensors (INT8/INT4 experts, BF16 attention)202.0download—————other or no referenceL36 @ 6c077c3
tcclaviger/MXFP4-FP8-GPTQ publisher comparison buildold static MXFP4 buildMXFP4+FP8—.063893.26%4.603——vs BF16L24 @ 6fe3e74
tcclaviger/MXFP4-FP8-GPTQ publisher comparison buildfirst activation-aware build (AA1)MXFP4+FP8—0.08693.0%———vs BF16L134 @ 6fe3e74
TheDrainFlorist/VQ-3.2bpw publisher comparison buildaffine q8 (ours)MLX191.1text weights (excl. vision, MTP)22.4 mnats94.8%2.367——vs BF16L179 @ 33e10e6
TheDrainFlorist/VQ-3.2bpw publisher comparison buildaffine q6 (ours)MLX147.1text weights (excl. vision, MTP)39.9 mnats93.4%2.276——vs BF16L178 @ 33e10e6
TheDrainFlorist/VQ-3.2bpw publisher comparison buildaffine q5 (ours)MLX124.5text weights (excl. vision, MTP)97.5 mnats89.0%2.549——vs BF16L176 @ 33e10e6
TheDrainFlorist/VQ-3.2bpw publisher comparison buildVQ-5.5bpwMLX119.8text weights (excl. vision, MTP)30.9 mnats93.7%2.244——vs BF16L177 @ 33e10e6
TheDrainFlorist/VQ-3.2bpw publisher comparison buildaffine q4 (ours)MLX103.1text weights (excl. vision, MTP)298.2 mnats77.8%3.912——vs BF16L174 @ 33e10e6
TheDrainFlorist/VQ-3.2bpw publisher comparison buildVQ-4.4bpwMLX100.5text weights (excl. vision, MTP)47.7 mnats92.2%2.323——vs BF16L175 @ 33e10e6
TheDrainFlorist/VQ-3.2bpw publisher comparison buildaffine q3 (ours)MLX80.53text weights (excl. vision, MTP)1312.4 mnats49.5%14.054——vs BF16L171 @ 33e10e6
TheDrainFlorist/VQ-3.2bpw publisher comparison buildVQ-2.1bpwMLX47.78text weights (excl. vision, MTP)652.4 mnats72.3%3.559——vs BF16L172 @ 33e10e6
TheDrainFlorist/VQ-4.4bpw publisher comparison buildaffine q8 (ours)MLX191.1text weights (excl. vision, MTP)22.4 mnats94.8%2.367——vs BF16L188 @ c13ef42
TheDrainFlorist/VQ-4.4bpw publisher comparison buildaffine q6 (ours)MLX147.1text weights (excl. vision, MTP)39.9 mnats93.4%2.276——vs BF16L187 @ c13ef42
TheDrainFlorist/VQ-4.4bpw publisher comparison buildaffine q5 (ours)MLX124.5text weights (excl. vision, MTP)97.5 mnats89.0%2.549——vs BF16L185 @ c13ef42
TheDrainFlorist/VQ-4.4bpw publisher comparison buildVQ-5.5bpwMLX119.8text weights (excl. vision, MTP)30.9 mnats93.7%2.244——vs BF16L186 @ c13ef42
TheDrainFlorist/VQ-4.4bpw publisher comparison buildaffine q4 (ours)MLX103.1text weights (excl. vision, MTP)298.2 mnats77.8%3.912——vs BF16L183 @ c13ef42
TheDrainFlorist/VQ-4.4bpw publisher comparison buildaffine q3 (ours)MLX80.53text weights (excl. vision, MTP)1312.4 mnats49.5%14.054——vs BF16L180 @ c13ef42
TheDrainFlorist/VQ-4.4bpw publisher comparison buildVQ-3.2bpwMLX73.77text weights (excl. vision, MTP)147.3 mnats86.7%2.300——vs BF16L182 @ c13ef42
TheDrainFlorist/VQ-5.5bpw publisher comparison buildaffine q8 (ours)MLX191.1text weights (excl. vision, MTP)22.4 mnats94.8%2.367——vs BF16L165 @ 7740086
TheDrainFlorist/VQ-5.5bpw publisher comparison buildaffine q6 (ours)MLX147.1text weights (excl. vision, MTP)39.9 mnats93.4%2.276——vs BF16L164 @ 7740086
TheDrainFlorist/VQ-5.5bpw publisher comparison buildaffine q5 (ours)MLX124.5text weights (excl. vision, MTP)97.5 mnats89.0%2.549——vs BF16L162 @ 7740086
TheDrainFlorist/VQ-5.5bpw publisher comparison buildaffine q4 (ours)MLX103.1text weights (excl. vision, MTP)298.2 mnats77.8%3.912——vs BF16L160 @ 7740086
TheDrainFlorist/VQ-5.5bpw publisher comparison buildVQ-4.4bpwMLX100.5text weights (excl. vision, MTP)47.7 mnats92.2%2.323——vs BF16L161 @ 7740086
TheDrainFlorist/VQ-5.5bpw publisher comparison buildaffine q3 (ours)MLX80.53text weights (excl. vision, MTP)1312.4 mnats49.5%14.054——vs BF16L157 @ 7740086
TheDrainFlorist/VQ-5.5bpw publisher comparison buildVQ-3.2bpwMLX73.77text weights (excl. vision, MTP)147.3 mnats86.7%2.300——vs BF16L159 @ 7740086
TheDrainFlorist/VQ-5.5bpw publisher comparison buildVQ-2.1bpwMLX47.78text weights (excl. vision, MTP)652.4 mnats72.3%3.559——vs BF16L158 @ 7740086
turboderp/exl3 measured by lackonendesEXL3 3.05 bpwEXL350.79trunk only————kl:code 0.089; kl:plain-English 0.057vs BF16L124 @ 257599e
turboderp/exl3 measured by JRamirez-UAB2.05 bpw EXL3EXL3—————kld public/internal 0.060 / 0.069vs BF16L119 @ 9eb541a
turboderp/exl3 measured by scorpoon3.05bpw h5EXL3—0.01730—1.4071+1.63%—vs BF16L63 @ 8165296
turboderp/exl3 measured by scorpoon2.05bpw h4EXL3—0.06840—1.4879+7.46%—vs BF16L64 @ 8165296
turboderp/exl3 measured by Vemloir4.05bpw_h6_ng6EXL3—————kl:Nemotron 0.00909; kl:multilingual 0.00655; ppl:Nemotron 1.9316; ppl:multilingual 2.3443vs BF16L40 @ c9e1197
txgsync/Dynamic-oQ5e-BF16-PLE-mtp publisher comparison buildOriginal Mac proxy calibration buildMLX——98.1079%1.28399468—kl_vs_bf16 0.01682755vs BF16L93 @ b62713a
UD-IQ4_XS GGUF on llama.cpp (publisher not named on card) measured by matteiuspiUD-IQ4_XSGGUF—————GPQA Diamond (AISBench Answer: X) 140/198 (70.71%); gpqa_final_response_correct 120/198 (60.61%); gpqa_hit_output_limit 74/198other or no referenceL187 @ d32978c
umixer/oQ4e-mix6-mtp publisher comparison buildoQ4e-mix6 v1MLX—————MMLU-Pro accuracy on finished 88.6%; MATHQA accuracy on finished 92.3%; MBPP accuracy on finished 97.3%other or no referenceL144 @ 73d2cd0
unnamed community 4-bit pack (not identified on card) measured by LOKSUN-1584-bit——87.82%——top1_delta_pts −2.42vs BF16L87 @ a6692a8
unsloth/GGUF measured by sirfyynUD-Q6_K_XLGGUF169.2download————kld:E 0.0116 ± 0.0008; kld:W 0.0164 ± 0.0005; kld:K 0.0363 ± 0.0012; top1_pct:E/W/K 98.5 / 95.1 / 93.0vs BF16L215 @ 1b64f8e
unsloth/GGUF quoted by Wladastic, not measuredUD-Q5_K_XLGGUF158.3download—————no quality claimL31 @ ba82efc
unsloth/GGUF measured by agentionaiUD-Q4_K_XLGGUF111.3download0.077887.21 %——kld:code 0.0192; top1:code 97.36 %reference unclearL71 @ 89f38ee
unsloth/GGUF measured by yamz-labsrepo not named on cardUD-Q4_K_XL (run in Gufo v0.8.0)GGUF111.3download0.0156 kl95.36 %——kl_run_to_run 0.0000other or no referenceL85 @ 779f63f
unsloth/GGUF measured by bitlamasUD-Q4_K_XL (reference)GGUF111.3download——2.9167 ± 0.0365——other or no referenceL48 @ 408a8e6
unsloth/GGUF quoted by Wladastic, not measuredUD-Q4_K_XLGGUF111.3download—————no quality claimL27 @ ba82efc
unsloth/GGUF measured by nazeshinjiteUD-Q5_K_XLGGUF103.8resident————nll 0.2831; greedy_first_token 88/100; nll:long 0.1209; dnll −0.0011 (−0.0032..+0.0010)other or no referenceL108 @ 9374dd4
unsloth/GGUF measured by sirfyynUD-IQ4_XSGGUF93.70download————kld:E 0.049 ± 0.003; kld:W 0.046 ± 0.001; kld:K 0.122 ± 0.004; top1_pct E/W/K 96.8 / 91.7 / 88.6vs BF16L84 @ 6bbaae7
unsloth/GGUF measured by TylogiUD-IQ4_XSGGUF93.70download0.08363 kld89.55%———reference not statedL91 @ 3cdfba4
unsloth/GGUF measured by sirfyynUD-IQ4_XSGGUF93.70download————kld:E 0.049 ± 0.003; kld:W 0.046 ± 0.001; kld:K 0.122 ± 0.004; top1_pct:E/W/K 96.8 / 91.7 / 88.6vs BF16L217 @ 1b64f8e
unsloth/GGUF measured by yamz-labsrepo not named on cardUD-IQ4_XS (run in Strata v0.1.40)GGUF93.70download0.0316 kl93.64 %——kl_run_to_run 0.0074other or no referenceL84 @ 779f63f
unsloth/GGUF measured by bitlamasUD-IQ4_XSGGUF93.70download—90.75 %2.998 ± 0.038—kld_vs_UD-Q4_K_XL 0.089other or no referenceL50 @ 408a8e6
unsloth/GGUF measured by drluotoUD-IQ4_XSGGUF93.70download——4.7393 ± 0.0291——other or no referenceL101 @ d69cf60
unsloth/GGUF measured by CyroniusUD-Q3_K_XLGGUF90.00download——2.40—tools .900; GSM8K-15 1.000; MMLU-30 .833other or no referenceL49 @ 2b7b5c7
unsloth/GGUF measured by TylogiUD-Q3_K_XLGGUF90.00download0.10650 kld88.31%———reference not statedL90 @ 3cdfba4
unsloth/GGUF measured by FINAL-BenchUD-IQ4_XSGGUF88.00unknown——6.013 ± 0.103——other or no referenceL96 @ cb1c47a
unsloth/GGUF measured by Maximilian228UD-Q4_K_XLGGUF82.68RAM in use—94.1%———vs a Q8_0L35 @ 379ff0b
unsloth/GGUF measured by Maximilian228UD-Q4_K_XLGGUF82.68RAM in use—94.1%———other or no referenceL42 @ 9ac833e
unsloth/GGUF measured by Maximilian228UD-Q4_K_XLGGUF82.68RAM in use—94.1%———vs a Q8_0L39 @ d8fc7df
unsloth/GGUF measured by nazeshinjiteUD-Q4_K_XLGGUF82.57resident————nll 0.2861; greedy_first_token 86/100; nll:long 0.1229; dnll +0.0019 (+0.0005..+0.0034)other or no referenceL110 @ 9374dd4
unsloth/GGUF measured by sirfyynUD-IQ3_XXSGGUF82.00download————kld:E 0.091 ± 0.005; kld:W 0.091 ± 0.003; kld:K 0.229 ± 0.007; top1_pct E/W/K 95.5 / 88.3 / 84.6vs BF16L88 @ 6bbaae7
unsloth/GGUF measured by TylogiUD-IQ3_XXSGGUF82.00download0.16512 kld85.41%———reference not statedL89 @ 3cdfba4
unsloth/GGUF measured by sirfyynUD-IQ3_XXSGGUF82.00download————kld:E 0.091 ± 0.005; kld:W 0.091 ± 0.003; kld:K 0.229 ± 0.007; top1_pct:E/W/K 95.5 / 88.3 / 84.6vs BF16L223 @ 1b64f8e
unsloth/GGUF quoted by Wladastic, not measuredUD-IQ3_XXSGGUF82.00download—————no quality claimL90 @ ba82efc
unsloth/GGUF measured by TylogiUD-Q2_K_XLGGUF78.90download0.22461 kld82.72%———reference not statedL88 @ 3cdfba4
unsloth/GGUF measured by sirfyynUD-Q2_K_XLGGUF78.90download————kld:E 0.133 ± 0.007; kld:W 0.125 ± 0.003; kld:K 0.299 ± 0.008; top1_pct:E/W/K 94.4 / 85.5 / 81.7vs BF16L225 @ 1b64f8e
unsloth/GGUF measured by TylogiUD-IQ1_MGGUF74.50download0.31474 kld79.69%———reference not statedL87 @ 3cdfba4
unsloth/GGUF measured by sirfyynUD-IQ1_MGGUF74.50download————kld:E 0.189 ± 0.008; kld:W 0.186 ± 0.004; kld:K 0.413 ± 0.011; top1_pct:E/W/K 92.6 / 83.4 / 79.5vs BF16L227 @ 1b64f8e
unsloth/GGUF measured by nibor1896UD-Q2_K_XLGGUF73.00download————oracle_kld 0.408 / 0.557; ten tasks greedy 2 Pass / 6 Partial / 2 Fail of 10other or no referenceL418 @ ce0ddcc
unsloth/GGUF measured by TylogiUD-IQ1_SGGUF72.50download0.39607 kld77.33%———reference not statedL86 @ 3cdfba4
unsloth/GGUF measured by Maximilian228UD-Q3_K_XLGGUF62.28RAM in use—90.7%———vs a Q8_0L38 @ 379ff0b
unsloth/GGUF measured by Maximilian228UD-Q3_K_XLGGUF62.28RAM in use—90.7%———other or no referenceL45 @ 9ac833e
unsloth/GGUF measured by Maximilian228UD-Q3_K_XLGGUF62.28RAM in use—90.7%———vs a Q8_0L42 @ d8fc7df
unsloth/GGUF measured by agentionaiUD-Q2_K_XLGGUF50.07GPU weights0.33476.61%———vs a Q8_0L62 @ aa0a0c3
unsloth/GGUF measured by turboderprepo not named on cardUD-IQ1_MGGUF—0.0804—1.5031——vs BF16L27 @ c1def42
unsloth/GGUF measured by turboderprepo not named on cardUD-Q2_K_XLGGUF—0.0533—1.4619——vs BF16L27 @ c1def42
unsloth/GGUF measured by turboderprepo not named on cardUD-IQ3_XXSGGUF—0.0349—1.4342——vs BF16L27 @ c1def42
unsloth/GGUF measured by turboderprepo not named on cardUD-IQ4_XSGGUF—0.0165—1.4017——vs BF16L27 @ c1def42
unsloth/GGUF measured by turboderprepo not named on cardUD-Q4_K_XLGGUF—0.0084—1.3926——vs BF16L27 @ c1def42
unsloth/GGUF measured by fuutottUD-Q4_K_XLGGUF—0.03693.9 %3.31 (× 1.005)—median_kld 0.007; MMLU (5-shot) 269 (89.7 %); GSM8K 92 (92.0 %); TruthfulQA MC1 192 (96.0 %); CMMLU (0-shot) 81 (81.0 %); bench_total 634 (90.6 %); HumanEval 157 (95.7 %); MBPP 429 (85.8 %)other or no referenceL111 @ 15ec9e8
unsloth/GGUF measured by AnonimousAUD-IQ1_M (unpruned)GGUF—————HumanEval-164 pass@1 95.7%; rare_name_probe 5/10; fabrication_rate 0%other or no referenceL38 @ 4bc31ec
unsloth/GGUF measured by AnonimousAUD-Q2_K_XL (unpruned)GGUF—————6-task battery (3 passes) 12 / 13 / 11; rare_name_probe 8/10other or no referenceL37 @ 4bc31ec
unsloth/GGUF measured by BoldingBuildsUD-IQ4_XS (stock)GGUF—————MATH-500 (200) correct 150; MATH-500 (200) empty final response 29; HumanEval pass@1 passed 151; HumanEval empty final response 11; StrongREJECT refusal 143 (95.3%); SimpleSafetyTests refusal 85; XSTest safe full compliance 88other or no referenceL33 @ bb56765
unsloth/GGUF measured by BoldingBuildsUD-IQ4_XS (stock) + --reasoning-budget 2048GGUF—————MATH-500 (200) correct 172; MATH-500 (200) empty final response 0other or no referenceL34 @ bb56765
unsloth/GGUF measured by nirecomUD-Q3_K_XLGGUF—————IFEval prompt_strict (thinking off) 0.8521; IFEval prompt_loose (thinking off) 0.8872; IFEval inst_strict (thinking off) 0.9017; IFEval inst_loose (thinking off) 0.9269; IFEval prompt_strict (thinking on) 0.9261; IFEval prompt_loose (thinking on) 0.9427; IFEval inst_strict (thinking on) 0.9472; IFEval inst_loose (thinking on) 0.9604other or no referenceL90 @ a158876
unsloth/GGUF measured by matteiuspirepo not named on cardUD-IQ4_XS (RTX 6000 PRO llama.cpp)GGUF—————GPQA Diamond first-5 correct 5 / 5; GPQA Diamond first-5 acc 100%; GPQA Diamond (198) 138/198, 69.70%vs BF16L338 @ 27e5cdc
Vontra 4-bit g32 (TensorFold; checkpoint not named) measured by beamster4-bit g3281.17GPU weights0.207485.01%———vs BF16L56 @ 76bd8f5
W4A16-RTN (checkpoint not named) measured by INCModel3W4A16-RTNINT4—————bench_avg acc_norm 0.8819vs BF16L92 @ ad4d80a
wtdcode/AWQ-W4A16 measured by primitive-aiAWQ INT4 g128AWQ—————tool_calling 76.0 (n=1)other or no referenceL215 @ e1b0da6
wtdcode/AWQ-W4A16 measured by primitive-aiAWQ INT4 g128AWQ—————tool_calling 76.0 (n=1)other or no referenceL252 @ 07915ee
zankich/W4A16-Merlin-INT4PLE-FP8KV publisher comparison buildsame checkpoint, default KV dtype (112 q subset)compressed-tensors W4A16 (GPTQ) + INT4 PLE + FP8 KV scales—————MMLU-Pro (112 q subset, default KV) 0.8839other or no referenceL102 @ ba8fff3

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.

6block/Qwen3.8-Flash-Next-GGUF

GGUF llama.cpp vs BF16

imatrix-calibrated tiers from BF16; protected tensors (router/SSM/HC F32, shared expert Q8_0, token_embd/output Q6_K, PLE table Q4_0); low tiers 'layered' (PLE Q4_0, savings spent on expert gate/up)

Reference
BF16 master (own run)
Eval data
wikitext-2 test, 12 chunks
Context
512
Tool
not stated (perplexity)
  • Tiny eval: 12 chunks x 512 tokens of wikitext-2 [L90]
  • Publisher says numbers are only comparable within this table [L91-92]
  • Claims layered tiers improve PPL ~14% vs the uniform tier at smaller size, but no uniform IQ2_XS/IQ1_M rows are shown [L52-53, L94-96]
  • Claim that PLE Q8_0 → Q4_0 'costs nothing' is stated without numbers [L50-51]
  • BF16 reference PPL 1.7164 on wikitext-2 is unusually low; other cards in this set report this architecture memorises wikitext via the n-gram table

Card snapshot 2d83ab731e53 · updated 2026-08-31

AesSedai/Qwen3.8-Flash-Next-GGUF

GGUF llama.cpp (mainline-compatible) reference not stated

MoE-aware mixtures: FFN up/gate/down quanted down, rest kept high; two families: Q8_0 PLE (n-gram) and Q4_0 PLE

Reference
not stated on card (repo KLD logs: BF16, wikitext-2 test, ctx 512)
Eval data
not stated on card
Tool
llama-perplexity --kl-divergence (per repo logs)
  • Column labelled 1-(Mean PPL(Q)/PPL(base)) but values are positive % increases (label/sign mismatch) [L13]
  • Evaluation method not stated on the card itself

Card snapshot d933da29a4d8 · updated 2026-10-03

agentionai/Qwen3.8-Flash-Next-AP-GGUF

GGUF llama.cpp mainline reference unclear

Agention Precision recipes (per-tensor distortion solver)

Reference
'the same source' (card); agention.ai says Q8_0
Eval data
held-out corpus (not wikitext — memorised by n-gram table)
Context
2048
Tool
llama-perplexity --kl-divergence
  • Reference model ambiguous: card says 'the same source'; Agention's site says Q8_0
  • Card reports wikitext inflates top-1 by 6.9–8.6 points for this architecture (n-gram table memorisation) [L95]

Card snapshot 89f38eeba9c7 · updated 2026-10-13

agentionai/Qwen3.8-Flash-Next-Gyro-GGUF

GGUF agentionai/llama.cpp fork (TQ types) / Strata vs a Q8_0

Gyro: rotated-expert ternary-ish TQ formats; calibration not published

Reference
unsloth Q8_0
Eval data
held-out 2026 technical writing + code, 60 chunks
Context
2048
Tool
llama.cpp fork KLD
  • Sizes are GPU-resident weights only (n-gram table on disk); Gyro-L download is 108 GB [L139]
  • Reference is unsloth Q8_0, not BF16
  • Same AP-Q4_K_XL file measures 0.107 here vs 0.0886 on the AP card — references differ

Card snapshot aa0a0c3e24f9 · updated 2026-10-16

agentionai/Qwen3.8-Flash-Next-ROCmFP4-FAST-imatrix-GGUF

GGUF agentionai/llama.cpp fork (ROCmFPx types) vs BF16

ROCmFP4 + imatrix (1540 chunks, 2 corpora)

Reference
unquantized PPL as reported in llama.cpp PR 27742 (not own run)
Eval data
wikitext-2 raw, 145 chunks
Context
2048
Tool
llama-perplexity
  • Uses wikitext PPL, which the same publisher's AP card says is memorised by this architecture
  • Baseline PPL borrowed from a llama.cpp PR, not measured alongside
  • AesSedai sizes quoted (107.38/117.13/135.38 GiB) don't match AesSedai's current card (105.98/114.99/131.98) — likely pre-10/03 re-quant

Card snapshot d37a14a0807a · updated 2026-10-11

airagrp/Qwen3.8-Flash-Next-mixed-iQ5-Q8-ple16-mlx-serve

MLX mlx-serve vs BF16

Experts iQ5 (imatrix affine g64) on layers 0–39, 8-bit on last 8; all other projections/heads/MTP 8-bit; n-gram table dense bf16 mmap; imatrix 162k tokens

Reference
bf16 next-token logits
Eval data
392 positions of held-out windows (row-matched n = 361–369)
Tool
mlx-serve tests/qwen38_flash_next_score.py (greedy, top-1024 logprobs)
  • Tiny eval: 392 positions (n = 361–369 row-matched) [L108-111]
  • Card says top-1 differences between bf16 and 8-bit heads are within noise; 4-bit embed scored higher top-1 [L123-124]
  • 188.1 GiB on disk includes 95.4 GiB n-gram table (non-resident); ~99.6 GB resident [L21, L25]
  • ddalcu iQ-MLX-4.7bpw re-measured here (top-1 89.8%); ddalcu's mixed-4-8bit card cites 88.6% top-1 for it (eval not stated there)

Card snapshot af5fc98deca9 · updated 2026-10-16

aixiaoma/Qwen3.8-Flash-Next-W4A16

compressed-tensors (W4A16) vLLM qwen4_exp dev build, --enable-expert-parallel mandatory other or no reference

RTN INT4 W4A16 g128 symmetric, no calibration: routed experts and QSA attention INT4; shared expert, MTP, PLE tables, router, indexer, GDN, embeddings, lm_head, vision BF16

Reference
Qwen3.8-27B INT8 W8A16 (a different model of the same family), not BF16 Flash-Next
Eval data
50-task deterministic greedy suite (arithmetic, sorting, positional retrieval, JSON extraction)
Tool
own suite
  • Reference is a different model (Qwen3.8-27B INT8), not BF16 Flash-Next [L88, L96]
  • Tiny eval: 50-task smoke suite; card itself calls it 'a 50-task smoke, not a full harness' [L90, L132]
  • Headline 'Quality ... matches the base-family reference' rests on that suite [L28]
  • 168 GB on disk includes 102 GB BF16 PLE tables; ~66 GB GPU-resident [L26]
  • 'First publicly available INT4' claim dated 2026-08-26 [L24]
  • Card's numbers match VnimanieAI/Qwen3.8-Flash-Next-W4A16's card (15 of 15 claim lines identical, detected automatically); treated as copied, not as an independent measurement

Card snapshot 75234a1d675c · updated 2026-08-27

amd/Qwen3.8-Flash-Next-Quark-MXFP4

MXFP4 vLLM ROCm vs BF16

Quark MXFP4 on all MoE experts incl. shared; attention bf16

Reference
BF16 (own run)
Eval data
gsm8k flexible-extract, 5-shot, chat template
Tool
lm-eval
  • Single saturated benchmark

Card snapshot b2091c3809eb · updated 2026-09-03

amd/Qwen3.8-Flash-Next-Quark-MXFP4-AttnFP8-PLEFP8

MXFP4+FP8 SGLang (PR #36601 + #39064), MI350X vs BF16

AMD-Quark: all MoE experts (routed, shared, MTP) OCP MXFP4; self/linear attention per-channel FP8 (dynamic per-token activations); PLE table per-tensor FP8

Reference
Qwen/Qwen3.8-Flash-Next (BF16), both with MTP
Eval data
gsm8k flexible-extract, 5-shot, chat template, thinking off
Tool
lm-evaluation-harness via SGLang
  • Single saturated benchmark [L94]
  • No size stated on card

Card snapshot bd6e49d9abfe · updated 2026-10-15

amd/Qwen3.8-Flash-Next-Quark-MXFP4-PLEFP8

MXFP4 SGLang (PR #36601 commit) on MI350/MI355 vs BF16

AMD-Quark: routed and shared MoE experts OCP MXFP4 (weights static, activations dynamic); PLE n-gram table per-tensor FP8; everything else bf16

Reference
Qwen/Qwen3.8-Flash-Next (own run)
Eval data
gsm8k flexible-extract, 5-shot, chat template, thinking disabled, greedy, max_gen_toks 2048
Tool
lm-evaluation-harness via SGLang chat-completions
  • Single saturated benchmark (GSM8K) [L93]
  • Text says 'SGLang framework' but the link points to docs.vllm.ai [L88]
  • Thinking disabled and generation capped at 2048 tokens [L111]
  • No size stated

Card snapshot 1ad36fa866c0 · updated 2026-10-01

aniolekx/Qwen3.8-Flash-Next-E9-MLX-8bit

MLX ENGINE (Swift/MLX) only; released mlx-lm/mlx-vlm untested other or no reference

8-bit affine group 64 for experts, projections, lm_head, embed_tokens, MTP head; n-gram table 8-bit group 32; routers, gates, norms, vision tower bf16 (routers/vision from Vontra conversion)

Reference
E9 itself has no direct BF16 comparison; predecessor E8h compared with a near-bf16 build (lm_head, HC mixes 8-bit g64, n-gram table 8-bit g32)
Eval data
teacher-forced NLL: 131,071 positions of a Polish novel [1]; technical corpus 16,383 positions of ENGINE source/Markdown and prose 32,767 positions [2]; 15 teacher-forced series (prose, private/public code, agent transcripts) [3]
Context
4096
Tool
engine nll (ENGINE build B44 for [3])
  • No direct E9-vs-BF16 measurement: the only near-bf16 comparison is for the predecessor E8h, and the reference build itself has 8-bit lm_head/HC mixes and an 8-bit n-gram table [L102, L294-298]
  • Predecessor scored lower NLL than the near-bf16 reference (1.79914 vs 1.80030); card notes the difference is not significant [L102]
  • All numbers come from an unpublished internal research log and have not been reproduced outside ENGINE [L96-98]
  • Prose NLL corpus is a single Polish novel [L294, L300]
  • Parity rows [4],[5] check runtime correctness, not quantization cost; card says so [L108-109]
  • 4-bit vs 8-bit code-loss attribution is DERIVED from per-file sums, not measured [L91, L304-305]

Card snapshot cfd4a0038aa2 · updated 2026-09-25

AnonimousA/Qwen3.8-Flash-Next-REAP-256-duo-GGUF

GGUF llama.cpp mainline (since 2026-08-27) modified weights other or no reference

REAP-style expert pruning 512 -> 256 per layer (binary copy of surviving experts) of Unsloth UD-Q3_K_XL; salience from llama-imatrix expert counts on code (164k) + agent/tool calling (180k) tokens

Reference
none (no unpruned or BF16 comparison on the battery)
Eval data
own machine-verified 6-task battery, 13 points; sampling temp 1.0, top_p 0.95, top_k 20
Tool
own harness
  • Pruned: half the routed experts removed; publisher marks it superseded and says K=256 prunes too deep [L21-22]
  • Battery is 6 tasks; card calls it 'a sharp instrument, not a broad one' [L204]
  • Earlier card version reported 13/13 from a single lucky run; 12/13 is now stated as typical [L145-150]
  • HumanEval 89.6% appears only in the superseded banner, with no method on this card [L22]
  • Base file is Unsloth's UD-Q3_K_XL (83.8 GiB) [L28, L71]

Card snapshot c80692368963 · updated 2026-09-06

AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF

GGUF llama.cpp mainline >= 2026-08-27 (+ expert-cache PR #27861 for measured numbers) modified weights other or no reference

REAP expert pruning 512 -> 320 per layer (salience from llama-imatrix over 525K-token multi-domain corpus), pruned by binary copy from unsloth UD-Q3_K_XL / UD-Q2_K_XL (no requantization)

Reference
unpruned unsloth quants (UD-Q2_K_XL, UD-IQ1_M), not BF16
Eval data
6-task machine-verified battery (3 passes); HumanEval-164 pass@1 (seed 42, effort low); rare-name probe (10 seeded gens); 12-question general-knowledge fabrication probe
Tool
llama.cpp + expert-cache PR #27861 build
  • Pruned: 192 of 512 experts per layer removed — not a faithful quant
  • Reference is unpruned 1-2 bit unsloth quants, not BF16 [L35-40]
  • Tiny probes: rare-name 10 generations, fabrication 12 questions; HumanEval single run [L38-40]
  • Card withdrew an earlier routing-mass retention claim (82.7% vs 75.3%) as circular [L25]
  • Throughput figures corrected twice; card states quality figures unaffected [L27-31, L304-306, L364-366]
  • Card calls the 94.5% vs 95.1% MTP HumanEval gap sampling noise [L99-101]
  • Card states all sizes are GiB (64.2 GiB = 68.96 GB) [L16]

Card snapshot 4bc31ecafb96 · updated 2026-09-09

apetersson/Qwen3.8-Flash-Next-Abliterated

GGUF llama.cpp (stock, qwen4exp support; verified df03399) modified weights other or no reference

Rank-one refusal-direction edit over 145 trunk writer tensors merged into the Q5_K_M trunk of apetersson/Qwen3.8-Flash-Next-GGUF (dequantize, add B·A, requantize to original type); BF16 n-gram table shard unchanged

Reference
stock Q5 trunk (apetersson/Qwen3.8-Flash-Next-GGUF), not BF16
Eval data
small non-thinking slices: IFEval 8, MATH-500 8, HumanEval+/MBPP+ 12, MMLU-Pro slice 8; sealed 6-cell sentinel refusal panel; bake fidelity on 7 sampled tensors
Tool
not stated (frozen scorers; different llama.cpp revisions for stock vs baked)
  • Weights modified: refusal-abliteration edit baked into the trunk [L21-29]
  • Very small slices (8-12 questions); card says they ran on different llama.cpp revisions and do not establish a general capability improvement [L43-47]
  • Denominators include unscorable cases (stock math 2, baked math 1, stock knowledge 1) [L43-44]
  • Bake fidelity: routed-expert IQ4_NL sample retained 0.34 of intended projection; card says samples do not establish fidelity for every tensor [L163-166]
  • Trunk shard size given as 78072.8 MiB (unit not GB/GiB) [L54]
  • Memory figures (95.5 GiB peak wired, 90.51 GiB peak) are measured on the stock base package, not this repo [L118-125]

Card snapshot 89bb1c93f195 · updated 2026-09-11

Ar3ntiz/qwen3.8_flash_next_side8

SUC1 SUC C++ / Metal runtime (suc1-generate, suc1-server) or SUC Studio; Apple Silicon only vs BF16

SUC (Semantic Universal Compression): routed experts 2-D vector codebook per tensor (VQ, G=2, K=256) with one-sided randomized Hadamard rotation and 0.5% activation-selected INT8 column residuals; attention/linear attention/shared expert/PLE k-v/LM head affine 8-bit g64; token embedding BF16; n-gram tables row-wise affine 4-bit; text-only (no vision, no MTP)

Reference
the original model
Eval data
16 × 512 held-out positions (KL, top-1); WikiText-2 test (PPL); answer-level 120 graded questions (20 per subset), greedy
Context
512
Tool
not stated (SUC runtime implied); single seed, Mac Studio M1 Ultra
  • KL eval over 16 × 512 held-out positions only; answer-level suite is 120 questions (20 per subset); card states single-seed measurements on one machine [L45, L59, L112]
  • Card's own comparison shows GGUF Q4 (ds4) with lower KL (0.061 vs 0.064) and higher answer total (99/120 vs 97/120) [L55-56, L63-64]
  • Comparison formats are not identified by repo; MLX affine 4-bit has no size [L56-57]
  • 4.35 bits per parameter stated to include every overhead; 73.82 GB fully resident at run time [L27-28]
  • Core suite of 30 graded smoke-test questions: 30/30 [L66]

Card snapshot ba49b2970580 · updated 2026-10-17

argyelan/Qwen3.8-Flash-Next-GGUF

GGUF llama.cpp with PR #27742 other or no reference

convert_hf_to_gguf from official FP8 checkpoint to Q8_0 (--fp8-as-q8), then llama-quantize --allow-requantize to IQ4_XS without imatrix; 51B n-gram table requantized to Q5_0 with custom chunked streaming requantizer

Reference
none (absolute scores, no BF16/FP8 comparison)
Eval data
HumanEval (164); GSM8K (200 problems)
Tool
llama.cpp server, 'greedy-ish defaults'; two-box DGX Spark + RTX 5090 over 10 GbE RPC
  • Absolute benchmark scores only; no BF16 or FP8 baseline measured on the card [L56-57]
  • GSM8K on a 200-problem subset; decoding described only as 'greedy-ish defaults' [L56-57]
  • Requantized from the official FP8 checkpoint (not BF16), no imatrix [L51]
  • Evals were run on a two-box RPC setup (DGX Spark + RTX 5090) [L38-40]
  • Card lists Q8_0 master (176 GB) as not yet uploaded [L26]

Card snapshot 22e87259f311 · updated 2026-08-26

arnomatic/Qwen38-AWQ-g32-ple8

AWQ (moe_wna16) + W8A16 PLE vLLM with AMD Qwen3.8-Flash-Next support + 4 patch files (2x Strix Halo, TP2+EP) other or no reference

Experts AWQ g32 asymmetric from leoncca/Qwen3.8-Flash-Next-AWQ-g32; PLE n-gram table baked to W8A16 g32 from the official FP8 checkpoint

Reference
reference W8A16 PLE table (aixiaoma/Qwen3.8-Flash-Next-W4A16 lineage)
Eval data
PLE table weights (weight-level comparison)
Tool
container/bake_awq_ple8.py
  • Only fidelity figures are weight-level for the PLE table (correlation 0.9996, max diff 2.29e-04 vs a reference W8A16 table); no output-level quality numbers [L30-31]
  • Asymmetric experts described as fixing word-form degradation seen with a symmetric RTN/g128 build ('reasoning ~3x shorter'), stated qualitatively without numbers [L25-27]
  • Expert quantization is leoncca's AWQ-g32, not re-done by this publisher [L23-24]
  • Memory ~68 GiB weights per rank (two ranks) [L38]

Card snapshot d40f5e9f8285 · updated 2026-09-09

AtomicChat/Qwen3.8-Flash-Next-GGUF

GGUF llama.cpp vs BF16

AD builds; imatrix on BF16, 4000 chunks @512, ~4.97M tokens

Reference
BF16 logits
Eval data
held-out neutral set
Context
4096
Tool
llama.cpp KLD
  • Headline '63% KLD reduction at the same file size' compares 84.9 GB vs 94.5 GB builds [L191, L83-84]
  • Card states 177B params in one place and 125B+51B+4B in another [L44, L258]

Card snapshot 1c1f33abc8e8 · updated 2026-10-13

AxionML/Qwen3.8-Flash-Next-NVFP4

NVFP4 SGLang (qwen4_exp) vs BF16

Unmodified mirror of RadixArk NVFP4 (ModelOpt v0.46.0, routed experts W4A4, calib 128 cnn_dailymail @512)

Reference
BF16 runs on an earlier base-model revision (RadixArk's)
Eval data
GSM8K full 1,319 (t=0.6, top_p=0.95); AIME 2026 30 problems x 8 (t=1.0)
Tool
not stated (scores reported by RadixArk)
  • Mirror: weights are an unmodified copy of RadixArk/Qwen3.8-Flash-Next-NVFP4 rev 7b71922; scores are RadixArk's, not re-measured [L25, L51]
  • BF16 baselines from an earlier base-model revision [L51]
  • Size given only as ~135 GB [L42]

Card snapshot 21d023c02aa4 · updated 2026-09-29

Baekpica/Qwen3.8-Flash-Next-Mixed-Quant-SSD-PLE-GGUF

GGUF (custom qwen4exp schema, ds4 SSD-PLE) Baekpica/ds4 dfm branch / ds4-dfm-rs (dedicated SSD-PLE loader; upstream llama.cpp/vLLM/SGLang compatibility not implied) other or no reference

Memory-hierarchy-aware mixed precision from pinned BF16: routed experts Q5_K/Q6_K/Q5_0 (Q6) or Q4_K/Q5_K/Q5_0 (Q5); MTP/always-active mostly Q8_0; 51.2B PLE n-gram table kept BF16 as non-resident SSD sidecar (optional FP8 sidecar extracted from Qwen FP8)

Reference
MQ-Q6 artifact (paired Q5 vs Q6 comparison), not BF16
Eval data
2,048 tokens from each of two fixed text fixtures (long-form essay; repetitive structured security fixture)
Tool
teacher-forced full-model comparison (ds4)
  • Only quality figures are a paired Q5-vs-Q6 NLL/PPL comparison on two 2,048-token fixtures; no BF16 reference [L1113-1121]
  • Card itself says the two slices are not a representative evaluation suite and do not establish general Q5 superiority; structured fixture is highly repetitive [L1123-1126]
  • Main GGUF sizes exclude the shared 95.3682 GiB BF16 PLE sidecar (full local setup 172.9240 GiB for Q5) [L166, L258, L263]
  • mem = resident backbone payload; bpw = resident backbone effective BPW (PLE excluded) [L250-257]
  • Repeated statements that serving checks do not establish full-model quality [L336-338, L372-374, L124-126]
  • Most of the card is runtime throughput/latency measurements on DGX Spark; some prefill rounds were developed on the Uncensored sibling artifact [L1144-1146]

Card snapshot f92da0134498 · updated 2026-09-13

batiai/Qwen3.8-Flash-Next-GGUF

GGUF llama.cpp mainline (PR #27742) other or no reference

Quantized from official BF16 with imatrix (4.90M chars: 20.13% Hangul, code, English wikitext; --chunks 200); PLE falls back to 32-block types (Q6_K rebuilt with Q5_1 PLE)

Reference
none (pass/fail functional checks, no BF16 comparison)
Eval data
Korean generation, trivial arithmetic, tool calls, coding, long-form Korean; external validation adds two arithmetic expressions
Context
8192
Tool
llama-server -ngl 0 -c 8192 -b 512 -ub 512 -np 1
  • Quality evidence is pass/fail functional checks and Hangul-share percentages; no perplexity, KLD or benchmark comparison with BF16 [L126-139]
  • External validation on the IQ4_XS files: two deterministic arithmetic prompts answered wrong (11,323 vs 11,289; 100 vs 200) [L107-114]
  • Length-constrained thinking test, 5 runs per quant: IQ4_XS 1/3, Q6_K 0/3 with constraint, both 2/2 without; card attributes it to the prompt, not quantization [L173-186]
  • Q6_K carries a Q5_1 PLE (38.40 GB), not the default Q8_0, due to the 50 GB per-file limit [L212-218]
  • Peak RSS 94.28 GiB for IQ4_XS on an external M2 Ultra at 8K context [L90]
  • Card's 'What BatiAI did differently' table says IQ3_XXS failed, while the PLE section says Korean checks passed at every tier including IQ3_XXS [L229-231, L306]

Card snapshot 1d0ec19ce9b0 · updated 2026-08-30

beamster/Qwen3.8-Flash-Next-Sushi-2.6bpw

EXL3 (sushi pack) sushi v1.0.4+ on Apple Silicon only vs BF16

Routed experts EXL3 2.6 bpw (publisher's own converter); dense 8-bit affine; n-gram table 4-bit affine g32 on SSD

Reference
bf16 checkpoint
Eval data
16 prompts x 512 tokens, scored to the first EOS (7186 positions)
Tool
sushi
  • Sizes are weights in GPU memory; the n-gram table (29.8 GiB here) stays on SSD and is excluded [L32, L44-45]
  • Each pack scored with the n-gram table it ships (4-bit, 4/5-bit oMLX, bf16), so KLD rows mix table precisions [L62-63]
  • Small eval: 16 prompts, 7186 scored positions [L43]
  • Other publishers' packs were run in sushi; oMLX n-gram tables repacked into sushi's file [L64-65]
  • LLMProbe Hard 50 is 50 questions, single run, no reference score [L71-72]

Card snapshot 0658b1533ff2 · updated 2026-10-07

beamster/Qwen3.8-Flash-Next-Sushi-2bpw

EXL3 (sushi pack) sushi >=1.0.4 only (Apple Silicon, macOS 26.2+) vs BF16

Routed experts EXL3 2 bpw (own converter); dense part 8-bit affine; n-gram table 4-bit affine g32 (mlx-serve's quantization) read from SSD

Reference
bf16 checkpoint
Eval data
16 prompts x 512 tokens, scored to first EOS (7186 positions)
Tool
sushi
  • Small eval: 16 prompts, 7186 scored positions [L43]
  • Sizes are weights in GPU memory excluding the n-gram table; disk is ~65 GiB (35.0 GiB weights + 29.8 GiB table) [L32, L44-45]
  • Sushi-2bpw scored with bf16 KV, other packs with kv8 [L43-44]
  • Packs carry different n-gram tables (4-bit, 4/5-bit, bf16), and other publishers' packs were re-run inside sushi [L61-63]
  • Same iQ-MLX 3.3bpw file: 86.28% top-1 here vs 85.6% on ddalcu's own card (different scorers)
  • LLMProbe Hard 50 is a 50-question set with no baseline [L69-70]

Card snapshot cb6c85e4f013 · updated 2026-10-07

beamster/Qwen3.8-Flash-Next-Sushi-3bpw

EXL3 (sushi pack) sushi v1.0.0+ (Apple Silicon, macOS 26.2+) only vs BF16

Routed experts EXL3 (own converter); dense 8-bit affine; n-gram table 4-bit affine g32 read from SSD

Reference
bf16 checkpoint
Eval data
16 prompts x 512 tokens, scored to the first EOS (7186 positions)
Tool
sushi (KV kv8; Sushi-3bpw/4bpw on a later build than the rest; Sushi-2bpw on v1.0.4 with bf16 KV)
  • Rows were not all run on the same build: Sushi-3bpw/4bpw on a later build, Sushi-2bpw on v1.0.4 with bf16 KV [L44]
  • Small eval: 16 prompts, 7186 scored positions [L43]
  • Sizes are GPU-resident weights; n-gram table excluded; disk is ~79 GiB [L32, L44-45]
  • Packs measured with different n-gram tables (4-bit, 4/5-bit, bf16) [L61-62]
  • showkieo/Qwen3.8-Flash-Next-Sushi-3bpw appears to republish this card's text (its download commands point to this publisher's repos) with different numbers for the same packs (Sushi-3bpw 0.1047 / 90.34%, Sushi-4bpw 0.0632); it is treated as a copy, not as an independent measurement
  • LLMProbe Hard 50 (82.0%) has no baseline [L69-71]

Card snapshot 76bd8f55df2b · updated 2026-10-07

beamster/Qwen3.8-Flash-Next-Sushi-4bpw

EXL3 (sushi pack) sushi v1.0.0+ on Apple Silicon only vs BF16

Routed experts EXL3 (publisher's own converter); dense 8-bit affine; n-gram table bf16 as released, read from SSD

Reference
bf16 checkpoint
Eval data
16 prompts x 512 tokens, scored to the first EOS (7186 positions)
Tool
sushi
  • Sizes are weights in GPU memory; the 95.4 GiB bf16 n-gram table stays on SSD and is excluded [L33, L47-48]
  • Rows were scored on different sushi builds and KV settings [L47]
  • Headline row uses the bf16 n-gram table while most compared packs use 4-bit tables; with the 4-bit table Sushi-4bpw reads 0.0654 [L64, L83]
  • Small eval: 16 prompts, 7186 scored positions [L46]
  • chatnerdie's card re-measures this pack at KLD 0.006137 on a different prompt set and settings and says the numbers are not comparable [chatnerdie L61, L77]

Card snapshot f25c0c4acdc4 · updated 2026-10-07

beginning-ai/Qwen3.8-Flash-Next-NVFP4

NVFP4 + FP8 (ModelOpt mixed) patched SGLang (beginning-ai fork, commit 93540334d2) vs BF16

ModelOpt 0.47.0: routed experts NVFP4 W4A4 g16, shared experts W4A16 NVFP4, attention FP8 128x128 blocks, PLE FP8 copied from Qwen FP8 release; activation scales max(own cnn_dailymail calib, RadixArk per-expert)

Reference
Qwen's reported BF16 DeepSWE score (113 tasks, high effort), quoted
Eval data
DeepSWE v1.1, fixed 30-task subset of 113, mini-swe-agent harness, reasoning effort medium, one attempt per task
Context
524288
Tool
patched SGLang, FP8 KV, MTP 3 steps / 4 draft tokens, 1x RTX PRO 6000
  • Baseline DeepSWE 58.7% is Qwen's reported score: quoted, not measured; different task count, effort and attempts [L93]
  • Card itself says the two scores are not directly comparable; 95% range about ±17 points [L95-99]
  • 'A 4-bit head changed the top token at about 20% of positions' refers to an unshipped variant (head kept BF16) [L68-70]

Card snapshot 9c8cb346a367 · updated 2026-10-10

bitlamas/Qwen3.8-Flash-Next-GGUF

GGUF llama.cpp mainline (b11681) / unsloth builds other or no reference

unsloth UD-Q4_K_XL with 48 ffn_down_exps replaced by IQ4_NL quantized from Qwen BF16 with unsloth imatrix

Reference
unsloth UD-Q4_K_XL (not BF16) for KLD/top-1; Qwen BF16 for down-projection error
Eval data
wikitext-2 test, 24 chunks of 2048 tokens
Context
2048
Tool
unsloth llama.cpp b10715 (perplexity / KLD)
  • KLD/top-1 reference is unsloth UD-Q4_K_XL, not BF16; publisher notes KLD against a quantized reference cannot rank quants with different rounding sources [L53, L128]
  • UD-IQ4_XS was scored against an earlier run of the reference (2.9117 ± 0.0363) on a different kernel path [L55]
  • Small eval: 24 x 2048 tokens of wikitext-2 [L53]
  • Stock builds since b11160 raise this file's KLD from 0.030 to 0.038 (int8 prompt matmul) [L74, L151]
  • 102.5 GB = 102,474,188,192 bytes (95.43 GiB) [L91]

Card snapshot 408a8e6ccece · updated 2026-10-17

BoldingBuilds/Qwen3.8-Flash-Next-Reasoning-Termination-Fix-GGUF

GGUF llama.cpp (PR #27742 branch commit 035e227; also loaded on mainline 5262471) modified weights other or no reference

unsloth UD-IQ4_XS (revision 8bdc6666) with one output.weight row (</think>, Q6_K) replaced by a fitted delta; all other bytes identical

Reference
stock unsloth UD-IQ4_XS (same files minus the edit) and stock + --reasoning-budget 2048
Eval data
MATH-500 first 200 after seeded shuffle (seed 7); HumanEval 164; StrongREJECT first 150, SimpleSafetyTests 100, XSTest safe first 100 (LLM judge)
Tool
llama-server, 16 parallel slots, file sampler defaults (temp 1.0, top-k 20, top-p 0.95, min-p 0.05), thinking xhigh
  • Weights modified: one output-layer row fitted (not a quantization); quant itself is unsloth's UD-IQ4_XS [L20-21, L96-97]
  • Comparison is to stock at a 4,096-token output budget only; larger budgets untested [L45, L199]
  • Card states p-values are exploratory and unadjusted; flag vs this model difference not significant (p = 0.39); HumanEval difference p = 0.29 [L37-42, L130]
  • Single seed, sampling at temperature 1.0 not bit-reproducible; MATH scorer strict, compare rows only [L161-168, L206]
  • Method developed on the 27B model using the same test sets [L111-112]
  • Size 'about 94 GB'; shards 10.9 MB + 49.84 GB + 43.84 GB [L58-60, L78]

Card snapshot bb5676573926 · updated 2026-09-29

btbtyler09/Qwen3.8-Flash-Next-GPTQ-4bit

GPTQ vLLM (main, Sep 2026+) / GPTQModel fork vs BF16

GPTQ W4 g32 symmetric, mse 2.0, RTN fallback for 7.6% of expert modules; routed+shared experts and full-attention projections INT4; GDN, indexer, routers, n-gram table, vision, MTP BF16; calib evol-codealpaca-v1 + C4, 2048 samples (~2.4M tokens)

Reference
BF16 (own run, identical windows)
Eval data
wikitext-2-raw-v1 test, 64 windows (32,768 scored tokens); also wikitext-2 4k ctx 8x1024 and codeparrot-clean-valid slice
Context
2048
Tool
transformers forward
  • mem = GPU-resident body (80 GB); 188 GB download includes the 102 GB BF16 n-gram table [L78, L80]
  • Headline PPL is on wikitext-2, which the card says the n-gram memory has close to memorized [L93]
  • Small eval sets: 32,768 wiki tokens; 4k-ctx suite only 8x1024 tokens [L86, L102]
  • Code slice described as 32x512 tokens in the table but '~65k scored tokens' in the footnote [L103, L105]

Card snapshot 9ae8d0e1989a · updated 2026-09-07

cc1244578151/myqwen3.8flashgsqrco

GGUF llama.cpp vs BF16

Card states GSQ + RCO per-tensor allocation; n-gram table fixed IQ4_NL (card text identical to ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF)

Reference
BF16 (as stated on the copied card)
Eval data
5 zero-shot tasks (arc_easy, arc_challenge, hellaswag, winogrande, piqa) + AIME25, GPQA-Diamond, LiveCodeBench v6
Tool
not stated
  • Card is byte-identical to ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF's card (incl. its 'Overview' and results); numbers are ISTA's, not measurements of this repo's files
  • Card itself says >100% recovery is within ~1 SE and means parity [L129]
  • BF16 baseline (LCB v6 87.43, GPQA 91.92) differs from Qwen's official 91.9 / 91.7 [L123]
  • Usage commands use a '<this-repo>' placeholder [L150, L175]
  • Card's numbers match ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF's card (48 of 48 claim lines identical, detected automatically); treated as copied, not as an independent measurement

Card snapshot a6341db0223d · updated 2026-10-08

chatnerdie/Qwen3.8-Flash-Next-EXL3-Sushi

EXL3 (sushi pack) sushi on Apple Silicon only vs BF16

Routed experts with open-source ExLlamaV3 v1.6.0 convert.py (-cb mcg, -sh scaled Hessian), calib 250 x 2,048 tokens generated by turboderp's 6.05 bpw EXL3; non-expert files copied from beamster Sushi-4bpw

Reference
BF16 checkpoint
Eval data
24 prompts (8 coding, 8 prose, 8 JSON extraction/tool calls); BF16 greedy 128-token replies, scored to first end-of-turn (2,516 positions)
Tool
sushi kld compare
  • Small eval: 24 prompts, 2,516 positions; card itself says not to over-read absolute values [L57, L116]
  • As-shipped KLD includes the YaRN x4 RoPE change (about 40% of 4-bit and 56% of 5-bit KLD per card) [L81]
  • Card says its numbers are not comparable with the 0.0592 on the Sushi-4bpw card (different prompts, reply length, KV, sushi build) [L77]
  • Sizes are weights in GPU memory; the 95.4 GiB bf16 n-gram table is not included and must be downloaded from Sushi-4bpw [L29]
  • Non-expert files are beamster Sushi-4bpw's, byte-identical [L18, L111]
  • 4bpw-sh misses one tool-calling prompt (4 greedy runs) with thinking off [L96, L103]
  • MMLU-Pro figures are from an earlier build with different calibration data [L83]

Card snapshot 5329132501fa · updated 2026-10-17

codelight-ai/Qwen3.8-Flash-Next-NVFP4-PLE-INT3

NVFP4 experts + INT3 PLE (compressed-tensors style) vLLM PR #53896 + repo's vllm-qwen4-exp-fixes.patch (not stock vLLM); GB10 / DGX Spark other or no reference

From RadixArk/Qwen3.8-Flash-Next-NVFP4: routed experts NVFP4 copied byte-identical; PLE n-gram tables re-quantized from Qwen BF16 to INT3 asymmetric per-row with per-row MSE-optimal clip search (3.20 bpw, no calibration); MTP head removed

Reference
BF16 PLE weights (weight-space rel-err ||Wq-W||_F/||W||_F and cosine)
Eval data
PLE table tensors (weight space; no end-to-end eval)
Tool
not stated
  • Only weight-space error figures for the PLE table; card states end-to-end quality has not been measured and INT3 is ~7x the error of the FP8 it replaces [L106-107]
  • Functional checks are a few short prompts and 4/4 tool-calling cases; long-form/agentic quality stated as unmeasured [L23-31]
  • MTP head removed: speculative decoding not available [L86-89]
  • Routed experts are RadixArk's NVFP4 tensors, byte-identical [L60, L81]
  • 99.3 GB on disk; 91.9 GiB resident [L64-65]

Card snapshot 59ecf4e3e99c · updated 2026-08-27

cygnal/Qwen3.8-Flash-Next-Heretic2-IQ4XS-NGQ4-GGUF

GGUF llama-flash-next fork of llama.cpp (qwen4exp branch) only modified weights other or no reference

Heretic v1.3.0+custom abliteration of BF16 -> Q8_0 intermediate -> IQ4_XS experts, Q8_0 dense, Q6_K output, BF16 token_embd; NGQ4: per-layer n-gram table IQ4_NL

Reference
none (no BF16 run; compared with other GGUF builds incl. a non-abliterated IQ4_XS NGQ4 build)
Eval data
HumanEval+ (EvalPlus), 164 problems
Tool
EvalPlus via llama-server
  • Abliterated (Heretic) weights: not a faithful quant of Qwen's model [L27, L119-121]
  • No BF16 reference; only comparisons among quantized builds [L83-91]
  • Single code benchmark (164 problems, thinking off, 1024-token cap), no run count or error bars; card itself calls the 79.3% vs 78.0% gap 'within noise' [L81, L128]
  • NGQ4 claimed to have 'no measurable accuracy impact through 32K context' with no numbers given [L38]
  • Q6_K builds' origin not stated; 129 GB size given only in prose [L86-87, L93]
  • Abliteration KL 0.0818 is Heretic's divergence from the original model, not quantization error [L123-126]

Card snapshot b7a350f3ba4c · updated 2026-09-19

Cyronius/Qwen3.8-Flash-Next-131B-A6B-GGUF

GGUF llama.cpp (b10673+) modified weights other or no reference

N-gram hash table pruned: keeps 1 of 8 heads per n-gram type (2 of 16), dropped heads point at a shared zero row; all other tensors byte-copied from unsloth UD-Q3_K_XL (or bartowski Q4_0)

Reference
unsloth UD-Q3_K_XL (the unpruned base quant, not BF16)
Eval data
40 agentic tool-calling cases; GSM8K-15 and MMLU-30 subsets (no-think); PPL on wikitext-2 test, 32x512-token chunks
Context
512
Tool
llama.cpp b10673, CUDA, A100
  • Weights modified: 14 of 16 n-gram hash heads removed — not a faithful quant [L20]
  • Wikitext PPL nearly doubles vs the base quant (4.66 vs 2.40); card says so [L50, L53, L79]
  • Tiny eval sets: 40 tool cases, GSM8K-15, MMLU-30 [L45, L80]
  • Tools 1.000 vs .900 is within a stated ±3-case run-to-run band; publisher reads it as parity [L53]
  • Q4_0 build not benchmarked; all numbers are from the Q3_K_XL build [L41, L82]

Card snapshot 2b7b5c71c597 · updated 2026-09-06

dabsLabs/Qwen3.8-Flash-Next-NVFP4

NVFP4 (compressed-tensors nvfp4-pack-quantized, W4A16) vLLM (vllm/vllm-openai:qwen38-flash-next image); end-to-end serving not yet done vs BF16

Weight-only NVFP4A16 on routed experts only, from BF16 (pinned f5d08274), with 4/6-optimal per-block scale selection; PLE, attention, shared experts, routers, embeddings, lm_head, vision, MTP kept BF16

Reference
BF16 source weights (weight-level relative error); cross-check vs Qwen official FP8 export
Eval data
all 73,728 expert projections; FP8 cross-check on 48 sampled experts
Tool
not stated (weight-space error)
  • Only weight-space error figures (per-tensor relative error vs BF16); no output-level quality numbers (KLD/PPL/benchmarks) [L30, L69-74]
  • Card states the checkpoint has not been served end-to-end yet [L37-40, L93-97]
  • Size ~174 GB; ~90 GB on-GPU weights with PLE CPU offload [L23, L42-43]

Card snapshot dc1c2354dc9e · updated 2026-08-29

ddalcu/Qwen3.8-Flash-Next-MLX-Serve-iQ-MLX-3.3bpw

MLX mlx-serve vs BF16

iQ-MLX: imatrix-measured per-layer 2/3/4-bit allocation; 482k-token calibration

Reference
bf16
Eval data
1,200 held-out next-token positions (agent, code, prose, math)
Tool
own scorer (qwen38_flash_next_score.py)
  • Download is 86 GB incl. 32 GB n-gram table; 52 GB is resident weights [L39]
  • mixed-4-8bit listed at 73 GB / 89.1% here, but its own card says 75 GB and top-1 87.9% vs its successor's 88.6% — the two cards disagree

Card snapshot 4fabeb5ee1ef · updated 2026-10-14

ddalcu/Qwen3.8-Flash-Next-MLX-Serve-iQ-MLX-4.7bpw

MLX (mlx-serve) mlx-serve vs BF16

iQ-MLX: every routed expert 4-bit g64 quantized with an importance matrix from bf16 (1.35 M tokens: agent transcripts, code incl. SWE-bench Lite, prose, math); rest 8-bit g64; n-gram 4-bit g32 mmapped

Reference
bf16 model's own logits
Eval data
290 held-out positions (agent, code, math, prose); KLD = KL(bf16 || pack) over bf16's top-1024 tokens
Tool
tests/qwen38_flash_next_score.py (mlx-serve repo)
  • Small eval: 290 held-out positions; card notes the KLD mean is dominated by a few flat prose positions and prefers the median [L54-57]
  • KLD computed over bf16's top-1024 tokens, not the full vocabulary [L55-56]
  • Prose top-1 is lower than the uncalibrated pack (81.2 vs 87.0) while code/math are higher [L61-62]
  • Download 107 GB incl. 32 GB n-gram table; 75 GB resident [L37, L91]
  • The sibling iQ-MLX-3.3bpw card lists mixed-4-8bit at 73 GB / top-1 89.1%, this card at 75.3 GB / 87.9 — the two cards differ (different held-out sets are possible: 1,200 vs 290 positions)

Card snapshot a0e89db090a3 · updated 2026-10-14

ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit

MLX mlx-serve reference not stated

4-bit group-64 experts, 8-bit dense

No evaluation method stated.

  • Superseded by the publisher's iQ-MLX-4.7bpw (KLD median 0.0052, top-1 88.6%) [L20-23]
  • Method for these numbers not stated on this card

Card snapshot 788c07a87256 · updated 2026-10-14

dealignai/Qwen3.8-Flash-Next-ABLITERATED-NVFP4

NVFP4 (ModelOpt FP4) SGLang (2x DGX Spark TP2) modified weights other or no reference

Abliteration (direct weight-level refusal removal) + NVFP4 W4A4 routed experts; attention, shared experts, PLE, vision, MTP higher precision

Reference
base model in identical harness (base precision not stated)
Eval data
MMLU 40 Q/subject x 57 subjects (2,280 Q); HarmBench-320 real-harm subset (240 behaviors), greedy temp 0
Tool
not stated
  • Abliterated: weights modified, not a faithful quant [L39-43]
  • MMLU subset of 40 questions per subject; per-subject deltas range from -12 to +15 points [L78-133]
  • Base-model run precision/quant not stated for the MMLU baseline [L69]
  • Table formatting: Overall row is appended to the 'world religions' row [L134]
  • No checkpoint size stated

Card snapshot be794b990578 · updated 2026-08-28

dealignai/Qwen3.8-Flash-Next-CYBERSECURITY-NVFP4

NVFP4 SGLang (2x DGX Spark TP2, modelopt_fp4) modified weights other or no reference

Refusal-removal weight edit ('CRACK') on an NVFP4 build; routed experts NVFP4 W4A4, attention/shared experts/PLE/vision/MTP higher precision

Reference
base model in NVFP4 (not BF16)
Eval data
MMLU-logit, 40 Q/subject (2,280 questions), greedy; hard multi-step question set (college math/physics/chemistry, formal logic; size not stated) at xhigh, greedy
Tool
not stated (same harness for both)
  • Abliterated / refusal-removed weights: not a faithful quant [L29, L43-51]
  • Reference is the base model in NVFP4, so the deltas measure the refusal removal, not quantization loss [L132]
  • MMLU subsampled to 40 questions per subject; hard-question set size not stated [L55, L130]
  • Per-subject MMLU swings up to -22.5 pp (machine_learning) against a -0.84 pp overall [L140, L190]
  • No absolute size given on card

Card snapshot 15ef113c7090 · updated 2026-09-04

Delgon/Qwen3.8-Flash-Next-PLE-GGUF

GGUF (PLE sidecar tables only) llama.cpp build with PLE sidecar support (--ple); not mainline, LM Studio or Ollama other or no reference

Standalone per_layer_token_embd (n-gram PLE) table in three precisions: BF16 byte copy of unsloth's BF16 GGUF table; Q8_0 converted from BF16; IQ4_NL byte copy of unsloth UD-Q4_K_XL table. Not a runnable model.

Reference
same model (main weights UD-Q4_K_XL) with the BF16 table
Eval data
held-out corpus written after the model's training cutoff; wikitext-2 as memorized-text control; 150 chunks
Context
512
Tool
llama-perplexity (KL divergence, top-1); Strix Halo (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)
  • Files are only the n-gram PLE table, not a runnable model; quality figures measure swapping this table into unsloth UD-Q4_K_XL, relative to the same model with the BF16 table [L21-23, L65-67]
  • IQ4_NL table is a byte copy of unsloth's UD-Q4_K_XL table; BF16 is a byte copy of unsloth's BF16 GGUF table [L37, L39]
  • Q8_0 table stated identical to AesSedai's Q8_0 table on 20,000 sampled rows [L38]
  • Card reports no measurable speed or MTP-acceptance difference between tables [L85-94]

Card snapshot 5de86c89b5b7 · updated 2026-10-07

DJLougen/Qwen3.8-Flash-Next-Mooney

GGUF (PQ2_0 ternary experts) buun-llama-cpp / DJLougen cuda.fast fork / prism-llama.cpp fork (stock llama.cpp cannot load) modified weights vs BF16

Mooney recipe: routed experts ternary (PQ2_0, 2.125 bpw) in a rotated basis, then recovered with distillation training; non-experts Q8_0; PLE table Q8_0 read lazily from SSD

Reference
BF16 original (own run on vLLM, 8xA100)
Eval data
own suite of 536 auto-verified tasks (unit tests, exact answers, schema and tool-call checks), run twice per model; needle retrieval 32k/64k/128k
Tool
own harness on vLLM
  • Experts recovered with distillation training after ternarization — weights changed beyond rounding [L86]
  • Quality numbers measured on vLLM 'with the same weight values', not on this GGUF; the GGUF was measured on DGX Spark for speed and memory only [L112]
  • Own task suite, no third-party benchmarks; card says so [L169]
  • mem ≈39 GiB is MemAvailable delta incl. KV/context (≈35 GiB weights); 92.0 GB download includes the 54.4 GB PLE table kept on SSD [L99-100, L108]
  • Long-text KL is truncated top-20 KL, not full-vocab [L127]

Card snapshot 84f867d3236f · updated 2026-10-07

donught/Qwen3.8-Flash-Next-NVFP4-dense

NVFP4 vLLM qwen38-flash-next image, TP=1, RTX PRO 6000 Blackwell other or no reference

ModelOpt 0.46.0 NVFP4_DEFAULT_CFG W4A4 on routed + shared experts, GDN projections and QSA o_proj; max calibration 512 x 1024 tokens (UltraChat 230, CodeFeedback 179, GSM8K-train 103)

Reference
none (no BF16 run; compared with primitive-ai mixed NVFP4-FP8 on the same box)
Eval data
GSM8K n=200, thinking on
Tool
vLLM (TP=1, PLE host offload, MTP 3 drafts)
  • No BF16 reference [L40-51]
  • Single GSM8K n=200 run; answers 19% longer than the comparison build [L50-51]
  • Calibration set includes 103 GSM8K-train samples; GSM8K eval split not stated [L64, L50]
  • mem = GPU-resident weights; PLE table served from host RAM [L38]

Card snapshot 8ce6fc792a80 · updated 2026-10-09

drluoto/Qwen3.8-Flash-Next-AgenticRequant-Q5K-GGUF

GGUF drluoto/llama.cpp strix-halo-vulkan branch (router quant needs patched build) other or no reference

Requant of unsloth UD-IQ4_XS: trunk IQ4_XS -> Q5_K, router F32 -> Q8_0 (--allow-requantize); experts and n-gram table unchanged

Reference
unsloth UD-IQ4_XS (source quant, not BF16)
Eval data
wikitext-2, 580 chunks
Context
512
Tool
llama-perplexity (same binary and flags)
  • No BF16 reference; compared only to its own source quant [L99-103]
  • Trunk requantized from IQ4_XS with --allow-requantize (rounded twice) [L167-169]
  • Publisher notes error bars overlap and PPL covers English prose only [L105-106]
  • Other quality gates (arithmetic at 24k, tool calling, vision) reported as pass with no numbers [L106-108]

Card snapshot d69cf601fafd · updated 2026-09-18

Dyluhn/qwen3.8-flash-next-exl3

radiance .rad (EXL3 experts) radiance 1.6.0 + radiance-exl3 plugins v1.2.0 (gfx1201) vs BF16

Routed experts (trunk and MTP) from turboderp EXL3 4.05bpw_h6_ng6; every other weight byte-identical to StillDeadcode/qwen3.8-next-flash-fp8-iq4r-moe (fp8 trunk, int8 lm_head, n-gram table, protected bf16 experts)

Reference
bf16 model's full-vocabulary log-probabilities
Eval data
ref64: 64 documents, 52,260 tokens, 52,132 scored positions (code, niche, prose, math, multilingual, agent traces, one 12K-token document)
Tool
radiance KL mode (each run twice)
  • Not an original quant: experts are turboderp's EXL3 4.05 bpw, rest copied from the radiance baseline container [L19-23]
  • Quantized PPL (4.0507) is below bf16 (4.0775) [L44]
  • ref64 eval is ~52K tokens [L34-35]
  • Speed figures from one machine with one card on PCIe 4.0 x4; card says this favours this container at high concurrency [L60-62, L137-138]

Card snapshot 473882c42552 · updated 2026-10-15

edougawa/Qwen3.8-Flash-Next-NVFP4-PLE-MMap-MTP-Spark

NVFP4 (ModelOpt W4A16_NVFP4) upstream vLLM 7c5dc571 + qwen38-nvfp4-mmap-ple plugin (DGX Spark, TP=1) vs BF16

ModelOpt W4A16_NVFP4 block 16 for GEMMs (experts, shared expert, QSA, GDN projections, PLE k/v proj, MTP linears, lm_head); BF16 for vision/embeddings/norms/routers; PLE table packed NVFP4 outside index, mmap

Reference
BF16 source tensors (weight-level cosine)
Eval data
PLE: 10000 random rows; body: all quantized tensors (per validation/body-nvfp4-validation.json)
Tool
publisher validators (validation/*.json)
  • Only fidelity figures are weight-level cosine similarities (PLE rows, body tensors); no output-level quality (KLD/PPL/benchmarks) [L84, L88]
  • Card states nothing has been booted on GB10; 32K instance test not run, long-context and memory envelope pending [L92-101]
  • Size: 69.9 GiB model shards plus 26.822 GiB external PLE table [L47-48]

Card snapshot c3e61234a968 · updated 2026-09-01

EliovpAI/Qwen3.8-Flash-Next-W3A8-Paiton-RDNA4

Paiton (W3A8) Paiton vLLM plugin container, 2x Radeon AI PRO R9700 only vs BF16

Own GPTQ: routed experts 8-level codebook ±{1,3,5,8} at 3.125 bpw in Hadamard-128 basis, fp8 activations; trunk E4M3 W8A16; int8 head; PLE E4M3; 2M-token calibration (best of two replicates)

Reference
bf16 model (emulation of the served arithmetic; served-KL gate on the running server)
Eval data
294,912-position evaluation corpus (170,884 text+assistant positions); long-v1: 44 held-out documents of 8K–32K tokens
Tool
own emulation + served-KL gate
Unit
full-vocabulary KL
  • Card says the served-KL gate passed at 0.0616 against a threshold of 0.0600 + 0.002 [L102-103]
  • Accuracy table numbers come from an emulation of the served arithmetic, not the running server [L100-101]
  • Publisher notes GPTQ replicate spread (0.0630 vs 0.0686) exceeds some table differences [L126-128]
  • No task benchmarks measured [L267-269]
  • Size 108.4 GiB includes 48.88 GiB PLE table held in host RAM [L246, L258]

Card snapshot e5cee0b46197 · updated 2026-10-14

EliovpAI/Qwen3.8-Flash-Next-W4A8-Paiton-RDNA4

Paiton W4A8 (INT4/INT3 hybrid experts) Paiton vLLM plugin, 2x Radeon AI PRO R9700 vs BF16

GPTQ (2M tokens): 410 most-used experts/layer INT4 in Hadamard-128 basis, 102 least-used 3-bit codebook; fp8 expert inputs; trunk W8A16 or INT6 g128; PLE E4M3

Reference
bf16 (original checkpoint on stock vLLM for tasks; bf16 on same tokens for PPL/KL)
Eval data
GSM8K 1,319, HumanEval 164, MMLU-Pro 1,400 subset (thinking on, temp 1.0, 16,384-token cap); PPL on wikitext-2 test and 22 unseen fineweb-edu docs (262,144 tokens); KL/top-1 on evaluation corpus 170,884 positions
Context
512 / 4096 (PPL)
Tool
served launcher for tasks; emulation of served arithmetic for PPL/KL
  • Limitations say task benchmarks (math, code, knowledge) were not measured for this release, contradicting the GSM8K/HumanEval/MMLU-Pro table [L235-236 vs L19-23, L42-46]
  • PPL/KL/top-1 measured on an emulation of the served arithmetic, not the served runtime [L62-63, L87-89]
  • MMLU-Pro on a 1,400-question subset; gap attributed to more replies hitting the 16,384-token cap (104 vs 79) [L52-55]
  • Card notes wikitext is partly memorised by the n-gram table, inflating differences there about 2x [L72-74]
  • 4bit-200k mode (8-bit trunk) not measured separately [L93]

Card snapshot 5e4a8dbee55d · updated 2026-10-14

evsinlb/Qwen3.8-Flash-Next-oQ4e-fp16-mtp

MLX (oMLX oQ4e) oMLX 0.6.4 / 0.7.0.dev2 other or no reference

oQ4e: 4-bit affine g64 with 5-bit boosts on sensitive tensors; PLE 5.00 bits/param; imatrix 1024 x 512 (oqe_code_multilingual); fp16 non-quant tensors

Reference
none
Eval data
3-scenario multi-turn agentic tool-loop battery (9 tool-call turns) + generation spot-checks
Tool
not stated
  • Only quality evidence is a 3-scenario agentic smoke battery and spot-checks; no reference comparison or standard benchmark [L91-93]
  • 70.00 GiB is always-resident weights; 29.82 GiB PLE table can be mmapped [L35-37]
  • Peak-memory figures in the throughput table are GB while sizes are GiB [L58-62, L37]

Card snapshot d2ec554e4697 · updated 2026-09-14

fabiogreter/Qwen3.8-Flash-Next-lily-q4-xl

lily qwen4_exp-affine-v1 (safetensors, engine-specific) lily-qwen3.8-flash-next (Metal, Apple Silicon) only other or no reference

Per-module affine: routed experts + GDN projections 4-bit g64; attention q/k/v/o, shared expert, LM head, embeddings 8-bit g64; draft head 4-bit; n-gram tables 4-bit g32; vision bf16 (converter --q4-xl)

Reference
sibling 4-bit checkpoint fabiogreter/Qwen3.8-Flash-Next-lily-q4 (behavioural comparison, no bf16 reference)
Eval data
one captured opencode turn with 56K tokens of context, 30 samples per checkpoint, thinking controls on
Context
56000
Tool
lily-qwen3.8-flash-next server
  • Only quality-related figure is a single-prompt behavioural count (samples that went on to a tool call: 23/30 vs 18/30 for the 4-bit sibling); card itself says 'one prompt, not a task-level evaluation' [L86-91]
  • No bf16 reference or KLD/PPL; card notes quantized output is not bit-identical to bf16 [L154-155]
  • Size 99.6 GiB across 55 safetensors files incl. draft head and vision; ~75 GB GPU-resident, 32 GB n-gram tables via page cache [L44, L97-99]

Card snapshot 0f94842e3682 · updated 2026-10-07

FenomAI/Qwen3.8-Flash-Next-GSQ-RCO-GGUF

GGUF llama.cpp / Ollama / LM Studio vs BF16

GSQ (Gumbel-Softmax scalar quant) + RCO per-tensor type allocation; n-gram shard fixed IQ4_NL

Reference
BF16 base model
Eval data
5 zero-shot tasks (arc_easy, arc_challenge, hellaswag, winogrande, piqa) + AIME25, GPQA-Diamond, LiveCodeBench v6
Tool
not stated
  • Results table and prose match ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF's card value for value; the card credits DASLab for the methods and compute and does not say who ran these evaluations or whether the files are ISTA-DASLab's [L63, L119-133, L226]
  • Usage commands still contain the placeholder '<this-repo>' [L150, L164, L175]
  • YAML license is apache-2.0 while the card says the weights inherit the base model's license [L6, L232]
  • Card itself says >100% recovery is within ~1 SE and means parity [L129]
  • Own BF16 baseline (LCB v6 87.43, GPQA 91.92) differs from Qwen's official 91.9 / 91.7
  • mem = shard 1 (must be resident); 28.8 GB n-gram shard can stay on disk [L89-94]
  • Card's numbers match ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF's card (48 of 48 claim lines identical, detected automatically); treated as copied, not as an independent measurement

Card snapshot 904afcb39146 · updated 2026-10-15

FINAL-Bench/POCKET-Qwen3.8-Flash-Next-GGUF

GGUF llama.cpp (stock) other or no reference

Q4_K_M from BF16; v2 proprietary bit allocation

Reference
none (compared to unsloth UD-IQ4_XS, not BF16)
Eval data
Korean Wikipedia 20231101.ko, 80 chunks (40,960 tokens)
Context
512
Tool
llama-perplexity
  • Lists unsloth UD-IQ4_XS as 88 GB; unsloth's docs give 93.7 GB (= 87.3 GiB), so the unit is probably GiB [L96]
  • No BF16 reference; publisher itself says all three are within error and claims no quality advantage [L98]

Card snapshot cb1c47ab3ebc · updated 2026-09-25

fuutott/Qwen3.8-Flash-Next-MLX-t5-imatrix-q3down-ple8

MLX (oMLX fork, ternary) Qwen48 oMLX fork only (Bonsai Metal kernel) other or no reference

Routed gate/up Bonsai T5 ternary g128 (~1.875 bpw, imatrix-weighted LSQ scales), routed down affine Q3 g128, PLE table Q8 g32 mmapped from SSD, shared experts Q8, embeddings/lm_head Q6, attention/DeltaNet Q5, rest Q4; Unsloth imatrix; MTP removed

Reference
Q8_0 teacher: lmstudio-community/Qwen3.8-Flash-Next-GGUF (PPL 3.297), not BF16
Eval data
KLD: wikitext-2 test, 40 windows x 2048 (40,920 scored positions); benchmarks: stratified subsets (seed 42) MMLU 300 5-shot, GSM8K 100, TruthfulQA MC1 200, CMMLU 100 0-shot; HumanEval 164, MBPP 500 (oMLX bundled)
Context
2048
Tool
llama.cpp --kl-divergence convention; MLX 0.32.2 on CUDA after repacking ternary to affine Q2
  • Reference is a Q8_0 GGUF (lmstudio-community), not BF16 [L101]
  • KLD on wikitext-2, English only; card notes Chinese/other languages/code not scored by KLD [L100, L153-154]
  • MLX checkpoints scored under MLX on CUDA after repacking ternary to affine Q2, not on the shipped Metal kernel [L102-103]
  • UD-Q4_K_XL benchmark column served by LM Studio (llama.cpp), a different runtime [L122]
  • Publisher says this checkpoint and the all-affine Q2 build are a statistical tie on KLD [L113-115]
  • CMMLU regressed vs the previous ternary build (68 vs 78) due to format-compliance failures; card explains [L129, L146-152]
  • Sizes: 90.4 GiB on disk incl. 53.6 GiB SSD-mmapped PLE shards; ~35.6 GiB resident [L63-65]
  • Card front matter says Apache-2.0 license, whereas other cards give Qwen Community License 1.0 [L6, L231]

Card snapshot 15ec9e857b55 · updated 2026-09-09

GBP-DE/Qwen3.8-Flash-Next-oQ5e-mtp

MLX oMLX 0.6.4 (SSD N-gram Offload, Lightning MTP) other or no reference

oQ5e enhanced/imatrix from BF16; group size 64; sensitivity map from Jundot/Qwen3.8-Flash-Next-oQ4e-mtp

Reference
none (head-to-head vs Jundot oQ4e only; no BF16 reference)
Eval data
five practical local prompts (8-person constraint puzzle, five-trigger state machine, Python debugging, four-tank state machine, 10-variable Boolean/SAT); one run each as presented
Tool
oMLX 0.6.4 on MacBook Pro M5 Max 128 GB, SSD N-gram Offload ON, Lightning MTP ON, TurboQuant KV/SpecPrefill/DFlash off
  • Quality evidence is pass/fail outcomes on five self-made prompts; card itself calls them 'practical local tests rather than a standardized academic evaluation suite' [L256]
  • Memory figures are total system memory including macOS, oMLX and other apps (e.g. Chrome), not model-only [L91]
  • No BF16 or KLD/PPL reference
  • Card states this earlier oQ5e is superseded by WLSRefit versions for new downloads [L27-38]

Card snapshot 387311fe0412 · updated 2026-10-13

glyd/Qwen3.8-Flash-Next-hummingbird

Glyd proprietary format Glyd engine (BSL 1.1, compiled) only other or no reference

Usage-based expert compression: most-routed experts keep most detail; precise ~2.7 bpw / fast ~2.1 bpw experts; non-expert weights (attention, embeddings, n-gram) in model/

Reference
none (absolute scores, no BF16 run on the card)
Eval data
Inspect AI GSM8K and MMLU-Pro as shipped, 150 questions each; one-prompt game task (Star Drift) x6
Tool
Inspect AI; greedy, thinking off, 3,072-token cap
  • No BF16 baseline on the card; scores are absolute on 150-question subsets [L85-91]
  • Sizes are model/ (37 GB) plus one expert folder (45 GB precise / 31 GB fast) [L55-59]
  • bpw figures are approximate ('about 2.7' / 'about 2.1') averages over experts only [L61-63]
  • Runs only on Glyd's own engine; not loadable in Transformers, vLLM or llama.cpp [L29-31]
  • Text only; vision encoder not supported [L95]

Card snapshot a101fa6e2ee0 · updated 2026-10-14

gpillon/Qwen3.8-Flash-Next-trellis-a25-ignis

trellis (EXL3/QTIP family, .ninfer) ignis (RTX 5090) only vs BF16

Routed experts trellis mul1 codebook, per-expert K in {2,2.5,3,4}, mean 2.50 bpw from calibration distortion; non-experts FP8 E4M3 per row; n-gram table Q4 g32

Reference
BF16 checkpoint
Eval data
teacher-forced 2,048-token windows across 11 domains (chat, code, prose, py, math, en, it, de, ja, zh, mmlu); window counts not stated; MMLU-Pro proxy 281 questions
Context
2048
Tool
ignis flash-next-converter (sidecar kld/mmlu fields)
  • Calibration text includes GSM8K test, MMLU-Pro questions, wikitext-103 test and ignis source — overlapping the math, mmlu, py and en eval domains [L169-177]
  • Per-domain window counts not stated; MMLU-Pro proxy is 281 questions [L68-69, L86]
  • FP8-only ablation scores above BF16 (74.7% vs 73.7%, McNemar p = 0.61) [L62-64]
  • Size 66.83 GiB excludes the 0.97 GiB MTP head file [L42-44]

Card snapshot f75e492c6d72 · updated 2026-10-12

GreenBitAI/Qwen3.8-Flash-Next-4bit-paged

MLX (gbx_lm paged containers) gbx_lm (single binary, Apple Silicon / DGX Spark); not mlx-lm readable other or no reference

Expert-paged repack of Vontra/Qwen3.8-Flash-Next-MLX-4bit; quantization unchanged; adds 1.52 GiB MTP draft head

Reference
source checkpoint Vontra/Qwen3.8-Flash-Next-MLX-4bit (equivalence check), not BF16
Eval data
5 prompts to 160 tokens + 1 longer
Tool
gbx_lm build-time gate
  • No quality claims for the quant vs BF16 (card read in full): only a build-time equivalence check against the source 4-bit checkpoint and MTP speed figures [L130-135, L206-222]
  • Not an original quant: repack of Vontra/Qwen3.8-Flash-Next-MLX-4bit [L25, L135]
  • Total 105.46 GiB incl. 1.52 GiB draft head; ~76 GB resident on 512 GB Mac Studio [L36, L66-67]

Card snapshot e98b855d64e1 · updated 2026-10-15

groxaxo/Qwen3.8-Flash-Next-exl3-unsloth-shaped-h8-mtp5

EXL3 ExLlamaV3 1.5.0 (TabbyAPI / exllamav3 server; tensor_parallel unsupported) other or no reference

Per-tensor 'Unsloth UD-IQ4_XS shape' EXL3 recipe, target 3.821 bpw: lm_head 6, attention/GDN 8, shared experts 6, middle expert up/gate 3-4; MTP 5, n-gram 6, vision 6 bpw; mul1 codebook; calib 250x2048

Reference
none (absolute correctness checks; no BF16 reference run)
Eval data
25-prompt quality battery: math 10, logic 5, code 5 (executed), misc 5, needle-in-haystack 2 (~20k, ~60k)
Tool
bench/bench_quality_v2.py, greedy, instruct mode, one pass per prompt
  • Small 25-prompt battery, single pass, no BF16 reference; card says these are absolute correctness checks, not a bpw-vs-BF16 delta [L209-211]
  • Earlier battery reported 21/25; card attributes the difference to a harness max_tokens artifact and reports a re-test [L186-195]
  • Card asserts the build 'should match or beat 3.50 HQ' with no comparative figures [L43-45]
  • Size ~104 GB = 65.2 GB shards + 39.0 GB n-gram table [L51]
  • Vision present but not exercised; all numbers text-only [L134-139]

Card snapshot 49725754024e · updated 2026-09-16

groxaxo/Qwen3.8-Flash-Next-Q2_0-H64-S4-GGUF

GGUF (Hadamard-rotated, patched runtime) patched llama.cpp (runtime/ in repo) other or no reference

Routed experts Q2_0 ternary g64 FP16 scales after Walsh-Hadamard blocks of 64 (seed 4); head/attention/shared experts Q8_0; n-gram Q4_0; no MTP or vision

Reference
none for prompts; weight reconstruction compared to direct Q2 on 26 weight samples
Eval data
27-prompt diagnostic set (thinking off) + retry of 6 misses with thinking on
Context
65536
Tool
3x RTX 3090, per-layer token table on CPU, MTP disabled
  • Card calls the prompts 'a useful smoke check, not a broad model benchmark' [L40]
  • Reconstruction-error figure is from the 26 samples used to choose the setting; card says it does not prove a general accuracy gain [L42]
  • ctx 65,536 is the configured allocation, not a checked quality window [L36]
  • No BF16 or KLD/PPL reference

Card snapshot 5eea3d52aecc · updated 2026-10-01

halt95/Qwen3.8-Flash-Next-W4A16-Merlin

W4A16 (INT4 AutoRound experts) + INT8 GDN + FP8 PLE patched vLLM from halt95/qwen38-flash-next-3090s (v2.5.2) vs BF16

Routed experts INT4 g128 from Intel AutoRound release; MTP experts INT4; GDN projections INT8 per channel; PLE FP8 (RadixArk); attention/norms/embeddings/vision BF16; FP8 KV with calibrated static scales

Reference
BF16 model
Eval data
GSM8K (200), structured output (80), 262K needle, divergence check — details not stated on the card
Tool
not stated (linked bench cards)
  • Quality summary is one line: 'gated against the BF16 model (divergence, 262K needle, GSM8K 198/200, structured output 80/80)'; no BF16 scores or divergence values on the card [L44-45]
  • Expert quantization is Intel's AutoRound release; PLE FP8 table from RadixArk [L117, L120]
  • Tested on one setup (4x RTX 3090) with the companion engine; stock vLLM not tested [L126, L130]

Card snapshot cec53e9f90e5 · updated 2026-10-12

HamboneLabs-AI/Qwen3.8-Flash-Next-uint3-g64

uint3 g64 (ModelOpt mixed) vLLM with included patches, DGX Spark GB10 other or no reference

Routed experts uint3 group 64 with bf16 scales, re-quantized from BF16; dense tail W4A16_NVFP4 and MTP NVFP4 carried from the source build; PLE bf16 mmap

Reference
'reference' (not identified on the card); token agreement 'against a lossless reference'
Eval data
GSM8K 200, MATH-500 200, HumanEval 164 (564 items, paired McNemar); token-level agreement set not described
Tool
not stated
  • Reference model for the accuracy table is not named; the card's other comparisons are against its 4.5-bit NVFP4 source build [L142, L121]
  • Publisher reports a significant regression: pooled −2.66 pp (p=0.0071), HumanEval −4.88 pp (p=0.0269) [L146-147]
  • Not all-from-BF16: dense tail and MTP drafter are the source build's NVFP4 tensors [L46-47]
  • 51.5 GiB excludes the 95.4 GiB bf16 PLE table (~147 GiB total download); mem = 67.48 GiB steady VRAM [L189-191, L227, L122]
  • Token-agreement eval set and size not described [L165]

Card snapshot 2da0f884d9f8 · updated 2026-08-29

IHaveNoClueAndIMustPost/Qwen3.8-Flash-Next-Simple-Q3_K

GGUF llama.cpp vs BF16

Simple mix on Q8_0 base: signal path Q8_0, gate/up Q3_K (edge blocks IQ4_NL), down IQ4_NL, n-gram table Q8_0

Reference
full BF16
Eval data
WikiText-2
Context
512
Tool
not stated
  • Total size stated as 114,245 MiB (core 60.9 GiB + 50.7 GiB n-gram table on disk); recorded size is the core excluding the table [L9]
  • Single PPL figure; perplexity tool and chunk count not stated [L19-20]
  • MTP not included in main files; separate standalone Q4_0 MTP GGUF added [L11-12]

Card snapshot 7ee3c667e10f · updated 2026-09-14

impulsai/CHImp-Alpha-v1.0

GGUF llama.cpp b10909 (unsloth CUDA-13 bundle) other or no reference

Requant from unsloth's Q8_0 conversion with unsloth imatrix (recipe htc2): expert input weights IQ4_XS 4.25 bpw (chosen to keep KLD vs Q8_0 < 0.05), expert output 4.5-bit, rest 6.6–8.5 bits or source precision; text only

Reference
Q8_0 conversion (175.3 GiB), not BF16
Eval data
wikitext-2 test, 64 chunks x 2048 tokens (131,072 tokens); EvalPlus HumanEval/MBPP, Terminal-Bench Core 0.1.1 (80 tasks), BFCL v4 (partial categories)
Context
2048
Tool
llama.cpp (perplexity/KLD); EvalPlus 0.3.1; terminal-bench 0.2.18; bfcl-eval 2026.3.23
  • Reference is Q8_0, not BF16; the build was requantized from that Q8_0, not from BF16 [L54, L261]
  • KLD/PPL on wikitext-2 [L133]
  • Task benchmarks have no reference run; card says the Q8_0 reference did not fit the machine [L72, L142-143]
  • BFCL partial: 11 of 13 single-turn and 1 of 4 multi-turn categories [L83]
  • Card says GGUF files not yet uploaded at this revision and build repo not public [L256-257, L268-269]
  • Card text says 177 billion weights; also 125B + 51B [L47-48]

Card snapshot 17645b9ea30f · updated 2026-09-24

INCModel3/Qwen3.8-Flash-Next-MXFP4-Mixed-CT-RTN-AutoRound

MXFP4 vLLM >= PR #55557 commit (0.29.1rc1.dev528 measured), TP=1 only vs BF16

AutoRound --model_free, scheme MXFP8 with mlp.experts at 4-bit mx_fp; llm_compressor format

Reference
same-box archived bf16 reference (4-task mean acc_norm only)
Eval data
gsm8k 5-shot chat template; piqa, mmlu, hellaswag (seed 42); niah_multiquery and ruler_qa_squad, 100 examples each at 131,072 tokens
Context
8192
Tool
lm-eval 0.4.13 on vLLM (thinking off, FP8 KV cache)
  • Repo named MXFP4 but the quantize command uses --scheme MXFP8 with only mlp.experts at 4-bit mx_fp [L38]
  • BF16 reference given only as a 4-task acc_norm mean, no per-task BF16 scores [L92]
  • Evaluated with FP8 KV cache [L44-45]
  • Mostly saturated short benchmarks; size not stated

Card snapshot ad4d80ad207f · updated 2026-09-22

INCModel3/Qwen3.8-Flash-Next-MXFP4-Mixed-FP8KV-CT-RTN-AutoRound

MXFP4-Mixed (llm_compressor) vLLM build supporting MXFP4-Mixed other or no reference

AutoRound model_free (RTN), experts mx_fp 4-bit via layer_config, base scheme MXFP8, static fp8 KV

Reference
none
Eval data
gsm8k 5-shot chat template (fewshot as multiturn); piqa, mmlu, hellaswag
Context
8192
Tool
lm-eval 0.4.13 on vLLM, thinking off
  • No baseline or recovery figure on the card [L26-31]
  • Quantize command uses --scheme MXFP8 with 4-bit mx_fp only for mlp.experts; eval path names 'mxfp-w8g32' [L38, L44]
  • Saturated/easy benchmarks only

Card snapshot fea777ab062d · updated 2026-09-24

inferencerlabs/Qwen3.8-Flash-Next-MLX-Q9

MLX Inferencer app (M3 Ultra); quantized with a modified MLX other or no reference

Q9 (9 bpw) quantization with a modified version of MLX; recipe not described

Reference
Base model ('predicting base tokens')
Eval data
not stated
Tool
not stated (metric definitions only: perplexity of base tokens, token accuracy, missed divergence)
  • Dataset, sample count, context length and tool for the table are not stated; only metric definitions are given [L43-45]
  • Table headline 'Q9-bit quant typically achieves near lossless accuracy' — card does not say whether the table was measured on this model [L25-38]
  • Ladder rows (Q3.5–Q8.5) are not linked to specific repos; Q4.5 and Q5.5 show identical Missed Divergence 17.28% [L33-34]
  • Only memory figure is ~189.6 GiB during multimodal inference on M3 Ultra; no file size stated [L17]

Card snapshot 627f6d688b43 · updated 2026-08-29

infosave/Qwen3.8-Flash-Next-cmf

CMF (cortiq) cortiq (Rust, Vulkan) >= 0.8.15 other or no reference

Quantized from bf16: q2tp experts 2-bit gate/up + 4-bit down, q4tp experts 4-bit; attention, embeddings and output head 8-bit; text only

Reference
none (no bf16 perplexity given)
Eval data
1,536 tokens of Rust source
Tool
not stated (cortiq)
  • Single small perplexity eval (1,536 tokens of Rust source) with no BF16 reference value [L69]
  • Vision encoder not included (text only) [L32, L198]
  • MTP files listed separately (1.08 GB / 1.50 GB) [L63-64]

Card snapshot d043fa6d77f7 · updated 2026-10-10

intel-ai/Qwen3.8-Flash-Next-MXFP4-BF16Attn-CT-AutoRound

compressed-tensors (MXFP4) vLLM >= 0.29.0 (TP=1) vs BF16

AutoRound 0.16.0 model-free RTN: routed experts MXFP4 W4A4 g32 UE8M0, dynamic activations; everything else BF16 incl. attention and PLE

Reference
Qwen/Qwen3.8-Flash-Next bf16 baseline (same table)
Eval data
GSM8K (chat template, few-shot multiturn), MMLU, PIQA, HellaSwag
Context
8192
Tool
lm-eval 0.4.13 + vLLM 0.29.1rc1.dev528, TP=1, bf16 KV, seed=42, enable_thinking=false
  • Baseline row lacks GSM8K flexible, PIQA acc_norm and HellaSwag acc_norm ('—') [L88]
  • Card does not state whether the bf16 baseline was run in the same harness
  • Saturated/easy benchmarks only, thinking disabled, max_model_len 8192 [L82-84]
  • NVIDIA comparison is size/format only, no quality numbers [L103-114]

Card snapshot 2855a0097727 · updated 2026-10-08

intel-ai/Qwen3.8-Flash-Next-MXFP4-BF16Attn-MXFP8MTP-CT-AutoRound

compressed-tensors (MXFP4/MXFP8) vLLM >= 0.29.0 (evaluated on 0.29.1rc1.dev528 nightly), TP=1, no patches vs BF16

AutoRound 0.16.0 --model_free (RTN, no calibration): main routed experts MXFP4 W4A4 g32 UE8M0 dynamic act; MTP routed experts MXFP8 W8A8 g32; attention, GDN, shared experts, PLE table, embeddings, head, vision BF16

Reference
bf16 Qwen/Qwen3.8-Flash-Next (same table; card does not state whether re-run or copied)
Eval data
GSM8K (chat template + few-shot multiturn), MMLU, PIQA, HellaSwag; MTP check GSM8K limit-100
Context
8192
Tool
lm-eval 0.4.13 + vLLM 0.29.1rc1.dev528, TP=1, bf16 KV, seed=42, max_num_seqs=64, enable_thinking=false
  • Saturated/easy benchmarks only (GSM8K, MMLU, PIQA, HellaSwag), thinking disabled, max_model_len 8192 [L78-80, L82-86]
  • BF16 baseline row has only one metric per benchmark (flexible / acc_norm shown as —) [L84]
  • Card states all deltas vs bf16-MTP sibling are <= ~0.3 σ; MTP weights are not loaded without speculative decoding [L88-90]
  • MTP check is GSM8K limit-100 only; acceptance/throughput not validated [L71-73]
  • Only size quoted for nvidia NVFP4 (132.7 GB), no quality comparison [L105-106]
  • Long-context (128k) not validated; validated B300, ctx <= 8k [L115-116]

Card snapshot 781020389a52 · updated 2026-10-08

intel-ai/Qwen3.8-Flash-Next-MXFP4-BF16Attn-MXFP8MTP-PLEMXFP8-CT-AutoRound

compressed-tensors (MXFP4/MXFP8) vLLM 0.29.1rc1.dev528 + mandatory PLE MXFP8 patch (patches/), TP=1 vs BF16

AutoRound 0.16.0 --model_free: main routed experts MXFP4 W4A4 g32 UE8M0 dynamic act; MTP routed experts MXFP8 W8A8 g32; PLE n-gram table MXFP8 (offline converter); attention, GDN, shared experts, embeddings, head, vision BF16

Reference
bf16 Qwen/Qwen3.8-Flash-Next (same table; card does not state whether re-run or copied)
Eval data
GSM8K (chat template + few-shot multiturn), MMLU, PIQA, HellaSwag
Context
8192
Tool
lm-eval 0.4.13 + vLLM 0.29.1rc1.dev528 with PLE patch, TP=1, bf16 KV, seed=42, max_num_seqs=64, enable_thinking=false
  • Saturated/easy benchmarks only (GSM8K, MMLU, PIQA, HellaSwag), thinking disabled, max_model_len 8192 [L93-101]
  • BF16 baseline row has only one metric per benchmark (flexible / acc_norm shown as —) [L99]
  • Card states deltas vs sibling are within ~1 σ and mixed in sign; claims MXFP8 PLE costs no measurable accuracy [L103-104]
  • Mandatory vLLM patch: unpatched vLLM silently misreads the PLE table and the model degenerates [L32-35, L127]
  • PLE MXFP8 SNR 53.6 dB is a weight-level fidelity figure [L30]
  • Validated only at context <= 8k; card advises long-context validation [L128-129]

Card snapshot 9008b69ed2eb · updated 2026-10-08

intel-ai/Qwen3.8-Flash-Next-MXFP4-BF16Attn-PLEMXFP8-CT-AutoRound

compressed-tensors (MXFP4 experts + MXFP8 PLE) vLLM 0.29.1rc1.dev528 + mandatory PLE MXFP8 patch (patches/) vs BF16

Intel AutoRound 0.16.0 --model_free: routed experts MXFP4 W4A4 g32 UE8M0 dynamic activations; PLE n-gram table converted offline to MXFP8 (E4M3 + UE8M0 g32); attention, GDN, shared experts, routers, MTP, embeddings, lm_head, vision BF16

Reference
Qwen/Qwen3.8-Flash-Next bf16 (own run)
Eval data
GSM8K, MMLU, PIQA, HellaSwag (lm-eval tasks)
Context
8192
Tool
lm-eval 0.4.13 + vLLM 0.29.1rc1.dev528 with patch, TP=1, bf16 KV, seed=42, enable_thinking=false
  • Four lm-eval tasks at max_model_len 8192 with thinking disabled; card says long-context (128k) behaviour of the MXFP8 PLE table is not validated [L95-97, L130-132]
  • Card's 'no measurable accuracy cost' claim compares against its own bf16-table sibling; deltas within ~1σ per card [L105-106]
  • Requires a vLLM patch; card states unpatched vLLM silently misreads the table and degenerates [L30-33, L129]
  • PLE SNR 53.6 dB is a weight-level fidelity figure [L28]
  • NVIDIA NVFP4 size 132.7 GB quoted for comparison, not a measured quality comparison [L26, L125]

Card snapshot b17da51a760d · updated 2026-10-07

intel-ai/Qwen3.8-Flash-Next-MXFP4-BF16SharedExperts-CT-AutoRound

compressed-tensors mixed-precision (MXFP4/MXFP8) vLLM >= 0.29.0 (evaluated on 0.29.1rc1.dev528), TP=1 vs BF16

AutoRound 0.16.0 --model_free (RTN, no calibration): routed experts MXFP4 W4A4 g32, attention and GDN MXFP8 W8A8 g32; shared experts, indexer, PLE, router, MTP, embeddings, lm_head, vision BF16

Reference
Qwen/Qwen3.8-Flash-Next bf16 baseline row in the same table (whether re-run under this harness not stated)
Eval data
GSM8K (chat template, few-shot multiturn), MMLU, PIQA, HellaSwag
Context
8192
Tool
lm-eval 0.4.13 + vLLM 0.29.1rc1.dev528, TP=1, bf16 KV, seed=42, enable_thinking=false
  • Baseline row gives only strict/acc values ('—' for flexible/acc_norm), so acc_norm figures have no BF16 comparison [L91]
  • Card states deltas vs the sibling variant are <= 0.6 sigma; standard errors themselves not printed [L95-97]
  • Size 180.0 GB dominated by the 102 GB BF16 PLE table [L39, L47-48]
  • NVIDIA NVFP4 entry is a size/design comparison only (132.7 GB), no quality numbers measured for it [L110-118]
  • TP>1 produces degenerate outputs on current vLLM's MXFP4 MoE path (card attributes to upstream) [L72-75]

Card snapshot a0ed54ef4582 · updated 2026-10-04

intel-ai/Qwen3.8-Flash-Next-MXFP4-Mixed-CT-AutoRound

compressed-tensors (MXFP4/MXFP8) vLLM nightly 0.29.1rc1.dev528, TP=1 only vs BF16

AutoRound model-free RTN (no calibration): routed experts MXFP4 W4A4 g32 E8M0; other Linear MXFP8 W8A8 g32; lm_head, embeddings, PLE table, MTP, routers, vision BF16

Reference
BF16 (own run, same protocol)
Eval data
GSM8K 5-shot (strict, chat template, greedy), MMLU, PIQA, HellaSwag 0-shot, full sample counts
Context
8192
Tool
lm-eval 0.4.13 + vLLM 0.29.1rc1.dev528
  • Saturated/easy academic benchmarks only, thinking off; card lists no long-context, agentic, code or thinking-on evaluation [L157-158]
  • Recovery headline '99.6 % of baseline' [L33]
  • Deltas stated to be within or near one SE (gsm8k ±0.53 pp) [L76-77]
  • Size 168 GiB is mostly the BF16 PLE table (102 GB) [L43, L55]

Card snapshot 2c55b10e3d88 · updated 2026-09-28

intel-ai/Qwen3.8-Flash-Next-MXFP8-CT-AutoRound

MXFP8 (compressed-tensors) vLLM (TP>=2, moe_backend=marlin) vs BF16

AutoRound 0.16.0 model-free RTN MXFP8 W8A8 g32 on all text-tower Linear incl. 73,728 routed experts; embeddings, lm_head, vision, MTP, PLE, hyper-connections, routers, in_proj_a/b, shared experts BF16

Reference
BF16 reference table supplied for the base checkpoint (not a local run, no stderr)
Eval data
gsm8k 5-shot chat (1,319), piqa 0-shot (1,838), mmlu 0-shot (14,042), hellaswag 0-shot (10,042)
Context
8192
Tool
lm-eval 0.4.13, vllm backend, TP=2 on 2x B300
  • BF16 row is a supplied reference table, not a matched local run; no BF16 stderr [L37-39, L185-188]
  • Publisher reports MMLU −1.37 pp (−4.8σ) and says not to describe it as lossless [L39-40, L172-177]
  • Same four tasks as Intel/Qwen3.8-Flash-Next-W4A16-AutoRound, whose BF16 row reads 0.9673 / 0.8651 vs 0.9674 / 0.8652 here [L163]
  • Saturated/easy benchmarks; thinking disabled [L118]
  • Shared experts left BF16 at export despite plan (deviation documented) [L190-197]

Card snapshot 8c4d7050d293 · updated 2026-09-21

Intel/Qwen3.8-Flash-Next-W4A16-AutoRound

AutoRound vLLM/transformers vs BF16

AutoRound W4A16 (routed experts)

Reference
BF16 (own run)
Eval data
GSM8K, MMLU, PIQA, HellaSwag
Tool
not stated
  • Four-task table (GSM8K, MMLU, PIQA, HellaSwag) against its own BF16 run; harness settings not given [L14-17]

Card snapshot 4c67bf686b7f · updated 2026-08-31

ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF

GGUF llama.cpp modified weights vs BF16

GSQ quantization + RCO expert pruning (256 of 512 experts kept), 3.5 bpw retained weights

Reference
own BF16 run
Eval data
SWE-bench Verified, LiveCodeBench v6
Tool
not stated
  • Pruned: half the routed experts removed — not a faithful quant
  • bpw 1.89 is amortised over original params (publisher explains) [L50]
  • Own BF16 LCB v6 87.43 vs official 91.9 — harness differs from Qwen's

Card snapshot 5348543e0147 · updated 2026-09-29

ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF

GGUF llama.cpp vs BF16

GSQ (Gumbel-Softmax scalar quant) + RCO per-layer type allocation; n-gram table fixed IQ4_NL

Reference
own BF16 run
Eval data
5 zero-shot tasks (arc_e, arc_c, hellaswag, winogrande, piqa) + AIME25, GPQA-Diamond, LCB v6
Tool
not stated
  • Publisher itself notes >100% recovery is within ~1 SE and means parity, not improvement [L129]
  • Own BF16 baseline (LCB v6 87.43, GPQA 91.92) differs from Qwen's official 91.9 / 91.7

Card snapshot ed59f92082b1 · updated 2026-09-29

ISTA-DASLab/Qwen3.8-Flash-Next-P48NVFP4-MoESQ

NVFP4 (W4A4) + paired-4:8 sparsity patched vLLM v0.30.0 (MoESQ), Blackwell SM100/SM120 only modified weights vs BF16

MoESQ: routed experts NVFP4 W4A4 with paired-4:8 structured sparsity, sparse storage; GPTQ init (512 samples) then masks and weights learned jointly for 10 epochs on 8,192 calibration sequences

Reference
BF16 (own run, same patched vLLM, TP4)
Eval data
OpenLLM v1 6 tasks (ARC-C 25-shot, HellaSwag 10-shot, MMLU 5-shot, TQA-MC2 0-shot, Winogrande 5-shot, GSM8K 5-shot chat) mean ± sd over 3 few-shot seeds; LiveCodeBench v6 175 problems, 10 seeds; SWE-bench Verified 500 instances, one run
Context
131072
Tool
lm-evaluation-harness, lighteval, mini-swe-agent 2.4.6
  • Weights modified beyond quantization: 50% structured sparsity with masks and weight values learned over 10 epochs [L35, L180-185]
  • LiveCodeBench: this model run with presence_penalty=1.5, BF16 at 0; without it this model scores 43.09 ± 1.33 at 65,536 tokens [L117-119, L131]
  • SWE-bench Verified: one run per model [L132]
  • At 131,072 tokens 19.8% of this model's LCB answers vs 1.4% of BF16's ran out of budget inside reasoning [L129-130]
  • Card says this model is more prone to repetition loops than BF16 on long reasoning [L74-76]
  • bpw 2.75 and 38.7 GiB cover routed experts only; 149.0 GiB on disk includes 95.4 GiB BF16 n-gram tables; mem = GPU-resident [L36-41, L147-149]

Card snapshot 80266c60ab7b · updated 2026-10-05

ivanfioravanti/Qwen3.8-Flash-Next-DS4-IQ2

GGUF ivanfioravanti/ds4-metal qwen3.8-flash-next branch, commit 5bd8796+ (Apple Metal); other runners not claimed vs BF16

IQ2_XXS imatrix gate/up experts; Q2_K imatrix down projections zero-padded 640->768 cols; Q8_0 dense/shared/output; BF16 embeddings; embedded MTP; external Q4_1 PLE sidecar (from Q4 repo); imatrix = pinned Unsloth imatrix (45 chunks x 18,432 tokens)

Reference
BF16 (Qwen rev de4b8e4d)
Eval data
99 BF16-reference continuations / 2,376 tokens; plus 12-question hard-smoke suite (2,048-token answer budget, temp 0, seed 123, 32K ctx)
Tool
not stated (DS4 harness; full results in qwen38-q2down.md)
  • Tiny eval: 2,376 tokens (99 continuations) and a 12-question smoke suite [L103, L110]
  • Main GGUF size excludes the REQUIRED 32.00 GB Q4_1 PLE sidecar; combined 76.81 GB on disk [L24-26, L48]
  • Pitched for 64 GB Macs [L21-22] but card says measurements 'do not establish a physical 64 GB fit' [L30]
  • Publisher itself says tests 'show extra quantization drift and do not establish equal reasoning quality' [L116-117]
  • Top-1 here (90.32%, 2,376-token continuations) is a different harness from the Q4 card's 96.6% (85 cases x 24 tokens) — not comparable across the two cards

Card snapshot b8b20398acfc · updated 2026-09-07

ivanfioravanti/Qwen3.8-Flash-Next-DS4-Q4

GGUF ivanfioravanti/ds4 qwen3.8-flash-next branch (Apple Metal) vs BF16

Q4_K imatrix routed gate/up, MXFP4 routed down, Q8_0 dense/shared/output, BF16/F32 embedding & control, embedded MTP; external Q4_1 PLE sidecar; imatrix = pinned Unsloth calibration (45 chunks of 18432)

Reference
exact-checkpoint local BF16
Eval data
100-case fixture, 24-token greedy continuations, top-20 logprobs; 85 tokenization-aligned cases
Tool
not stated on card (methodology in repo QWEN38_FLASH_NEXT.md / OUT/qwen38-eval/)
  • Tiny eval: 85 aligned cases x 24 greedy tokens (~2,000 tokens) [L124-126]
  • Main size excludes the REQUIRED 32.0 GB Q4_1 PLE sidecar (demand-paged, not resident) [L58, L106]
  • New Q4_K recipe (MAE 0.0431) is not better than the previous Q4_0-routed release (0.0430) — publisher says 'holding overall drift flat' [L129-131]
  • Hub sidebar quant label (Q4_1) refers to the PLE sidecar, not the experts [L50-51]
  • Unsloth imatrix revision/SHA on this card are malformed (39-char revision, 65-char SHA-256) and differ from the IQ2 card's 40/64-char values [L43-44 vs IQ2 L137-138]

Card snapshot 59a55fb819c8 · updated 2026-09-05

Jakevin/Qwen3.8-Flash-Next-ternary-mlx

MLX (mlx-ternary mixed) mlx-ternary + oMLX mlx-vlm build other or no reference

288 units: 117 ternary (~1.85 bits, scale per 64), 92 4-bit, 62 3-bit, 17 2-bit affine g64; started from official FP8 weights

Reference
official FP8 (not BF16)
Eval data
held-out text (size not stated); 216-question multiple-choice set (unnamed)
Tool
not stated
  • Reference is official FP8, not BF16 [L19, L30]
  • Held-out KL text size and MCQ set identity not stated [L30, L32]
  • Card itself says this mix answers fewer questions than ISTA 2-bit (149 vs 162) despite lower KL [L32]
  • Which ISTA 2-bit file was compared is not stated [L30]
  • Needs the official FP8 checkpoint at runtime for unquantized tensors; no size stated [L40]

Card snapshot e5f8a6bbb154 · updated 2026-10-13

jamesrogers/Qwen3.8-Flash-Next-MTP-MXFP4-GGUF

GGUF ik_llama.cpp main (MTP files); stock llama.cpp for -ngramBF16 only other or no reference

MXFP4 routed experts, attention/shared expert Q8_0, router/norms F32, n-gram table BF16 or Q8_0, MTP head appended (experts Q8_0); from official BF16

Reference
same quant with BF16 n-gram table (not the BF16 model)
Eval data
'a fixed corpus' (not named)
Tool
not stated (perplexity)
  • Only quality number is +0.12% PPL between the Q8_0 and BF16 n-gram-table variants, on an unnamed corpus [L54-55, L175]
  • Versus the reference conversion the card says only 'within noise', with no numbers [L174]
  • 0.93–0.99 figures are MTP draft acceptance (speed), not quality [L32]

Card snapshot 0a9ac88fa60e · updated 2026-09-03

JANGQ-AI/Qwen3.8-Flash-Next-JANGH4

MLX (JANGH) vMLX development build only (not in a released build) vs BF16

JANGH 4-bit experts (Hadamard-32 + closed-form codebook, GPTQ per expert), down proj 6-bit in 38 layers; attention/shared/lm_head 8-bit imatrix+AWQ; MTP 6-bit; n-gram table 8-bit g32 on SSD; calibration 604,179 tokens

Reference
full bf16 model run layer by layer from disk; 'ceiling' = bf16 with 0.001 relative noise on one tensor
Eval data
3 held-out sets: general (215 seqs, 11 domains), reasoning traces, tool conversations; 510,954 teacher-forced positions
Tool
own scorer; top-128 renormalized KL
  • KL is top-128 renormalized, not full-vocabulary [L66-67]
  • Comparison ceiling is a perturbed bf16 run (0.001 noise on one tensor), not bf16 itself [L60-63]
  • Runtime support not in a released vMLX build [L36-38]
  • mem = 72.3 GiB weights in memory (headline ~74 GiB); n-gram table 53.6 GiB read from SSD; 125.6 GiB on disk [L42, L135, L167]

Card snapshot 28cb56ceaec1 · updated 2026-10-04

jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4

GGUF (custom CIRU IU4 format) CIRU runtime (custom llama.cpp/ggml fork) on Strix Halo vs BF16

Routed experts 4.5175-bpw compact indexed affine (4-bit codes + scale/offset indices); rest mixed higher precision; PLE table SSD-paged

Reference
BF16 full-vocabulary reference
Eval data
2,048 next-token positions (data source not stated); task checks: 15-case hard tool set, HermesAgent-20 (two passes), thinking off
Tool
each stack's own runtime; Ryzen AI Max+ 395 / Radeon 8060S
  • Small objective panel: 2,048 positions; card says it does not establish long-context factual accuracy [L178, L196]
  • Rows compare complete weight+runtime stacks, not quants in isolation [L181-182]
  • Full quality panel and task suite were not rerun on the final v5.5.0 release build [L171-174]
  • In the card's own table v5.0 KL (0.050030) is higher than HaloBox, Gufo and AgentionAI AP [L186-191]
  • Halogen memory (34.47 GiB) is not comparable; card notes 78.16 GiB process PSS [L191, L193]
  • HumanEval 0-9 used for speed, not pass@1 [L207-208]
  • Historical v4 benchmark results come from earlier runtime/settings and are not comparable [L275-282]

Card snapshot f3f020e2545c · updated 2026-10-07

jcbtc/Qwen3.8-Flash-CIRU-STRIX-Orca

GGUF (CIRU compact expert format) CIRU v5.5.0 native runtime only (stock llama.cpp does not load it) modified weights other or no reference

Routed experts as 4-bit integer codes, groups of 16 sharing scale/offset table index (4.5175 bpw on routed experts); down projections fitted from Orca BF16 with Orca calibration activations; source is OrcaRouter's refusal-removed checkpoint

Reference
retained (original) Orca CIRU GGUF, not BF16 (card: full-precision Orca logit reference not available)
Eval data
sixteen fixed 512-token text windows, 2,048 tail positions scored
Context
512
Tool
not stated (CIRU runtime)
  • Source checkpoint is OrcaRouter's refusal-removed (uncensored) Qwen3.8 Flash [L6, L302, L316]
  • Reference is the publisher's earlier Orca GGUF, not BF16; card says old-to-new agreement is not BF16 agreement [L189-190]
  • Small eval: 2,048 scored positions, described by the card as a short-text diagnostic [L185, L188]
  • mem = GPU-visible unified memory after loading (94.19 / 87.31 GiB); 4.5175 bpw covers routed experts only [L116-117, L193]
  • Task scores (HumanEval, HermesAgent) are historical, on the original weights and earlier runtimes, not rerun on v5.0 [L281-290]

Card snapshot 139c9f5eb561 · updated 2026-10-07

jedisct1/Qwen3.8-Flash-Next-oQ4e-100K-MTP

MLX oMLX 0.6.3 other or no reference

oQ4e affine g64 with strict oQe imatrix (937 modules, 1,024x512); PLE table 4-bit g32; uncalibrated experts in base layers 0-1 kept Q8; lm_head, hyper-connection mixers, MTP fusion projections BF16

Reference
none for task tests (no BF16 run); Q8 vs Q4 reconstruction error measured on official BF16 rows
Eval data
Swival tool-calling suite: 50 positive cases + 10 adversarial guards; long-context retrieval markers at ~100K tokens
Context
100000
Tool
Swival via oMLX OpenAI-compatible API
  • Quality evidence is a functional tool-calling suite (50 positive + 10 guards) and needle-style retrieval, not KLD/PPL/standard benchmarks; no BF16 comparison [L146-151, L122-124]
  • Q8 'reduced normalized reconstruction error by 164.46 times compared with Q4' applies only to the affected uncalibrated expert rows [L41-42]
  • With MTP enabled one speculative run changed an exact tool argument (9/10 guards) [L148-150]
  • Recommended 100,000-token context on 128 GiB Mac; peak memory figures 87.84–99.96 GiB [L101-120]

Card snapshot a7b39c732acd · updated 2026-08-29

ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF

GGUF ik_llama.cpp cb860a9+ (CUDA validated; Metal/Vulkan/ARM/multi-GPU not validated) vs BF16

KT trellis quants: routed down IQ3_KT; gate/up IQ4_KT (v3 IQ4_KT) or IQ4_KT/IQ3_KT by block (v3 IQ3_KT); attention/SSM/HC/shared Q8_0; DSA indexer BF16; imatrix 700 chunks x 2048 agentic-coding corpus

Reference
BF16 GGUF converted from released weights
Eval data
two held-out texts (English; code = CPython + PostgreSQL source), 10 chunks x 2048 tokens each
Context
2048
Tool
ik_llama.cpp KLD (llama-perplexity implied, not named)
  • Small eval: 10 chunks x 2048 tokens per text (20,480 tokens each) [L106]
  • Replaced files measured with an earlier ik_llama.cpp build and earlier BF16 conversion [L116]; the '56%/44% lower KLD' headline [L25] compares across those runs (arithmetic checks out: 56.4%/55.3%, 43.7%/42.4%)
  • Eval chunk = 2048 = DSA indexer selection size, so sparse-attention selection is not exercised [L118]
  • 'Choosing a file' table gives rounded KLD (0.058/0.085/0.133) [L37-39]; precise values at L110-112
  • GPU-resident column (4.5 GiB) excludes routed experts and KV; not a disk size [L35, L45]
  • Calibration corpus is agentic-coding-weighted (45% technical English, 30% code) [L122]

Card snapshot 8fb1860c4e78 · updated 2026-10-15

ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF_v2

GGUF ik_llama.cpp cb860a9+ (CUDA; RTX 4070 + 64 GB RAM) vs BF16

IQ4_KT base; IQ3_KT for ffn_down_exps and ffn_down_shexp; IQ4_KT PLE table; Q8_0 token embeddings; imatrix 700 chunks x 2048 agentic-coding corpus

Reference
BF16 source (same references for all runs)
Eval data
English + code (CPython, PostgreSQL), 10 chunks x 2048 tokens each
Context
2048
Tool
not named
  • Superseded by v3 files in the main ji-farthing repo (lower KLD) [L22]
  • Small eval: 10 x 2048 tokens per text [L67]
  • Says all runs used 'the same BF16 references' [L67], but the main card says the previous IQ4_KT/IQ3_KT were measured against an earlier BF16 conversion and earlier build [main L116] — the two cards disagree
  • Quality runs used earlier file revisions 563ede2/8eb6566; current files are a lossless byte permutation [L32, L77]
  • KLD essentially equal to the previous IQ4_KT (0.132805 vs 0.132548 English) — gain is size, not quality [L72-75]

Card snapshot ca863d1404fa · updated 2026-10-15

jomcgi-org/Qwen3.8-Flash-Next-NVFP4-oominf

NVFP4 oominf (oom-inference) only vs BF16

RadixArk NVFP4 (ModelOpt 0.46.0, routed experts W4A4) re-laid out one record per expert; no weight changed

Reference
BF16 numbers quoted from RadixArk (earlier revision) and Qwen card (GPQA), not own runs
Eval data
GSM8K full 1,319 (t0.6, top-p 0.95, max 8,192, 1 sample); GPQA-Diamond first 118 of 198; temperature-0 paired runs on GSM8K 300, MMLU-Pro 200, RULER 8k/32k (104) and 95k (52)
Tool
oomeval via oominf serve (fp32 compute, k8v6 KV cache)
  • Same weights as RadixArk/Qwen3.8-Flash-Next-NVFP4; only the file layout differs [L68-72]
  • BF16 baselines are quoted (RadixArk earlier revision; Qwen card GPQA with unstated harness), not measured alongside [L98-102]
  • GPQA-Diamond partial: first 118 rows, not a random subset [L95-97]
  • MMLU-Pro at 4,096-token limit: 17.5% of answers cut off [L120-121]
  • Size: card says about 135 GB; files listed 73.1 + 51.2 + 10.1 GB [L151, L173-175]

Card snapshot 84010e0487b5 · updated 2026-10-09

Jon-Nielsen/Qwen3.8-Flash-Next-AWQ-INT4-FP8PLE

compressed-tensors AWQ INT4 + FP8 PLE wtdcode/vllm-backport fork (with ple_layer.py overlay) other or no reference

Repack of cyankiwi/Qwen3.8-Flash-Next-AWQ-INT4 (routed experts AWQ INT4 asym g32) with the n-gram PLE table converted BF16 -> FP8 e4m3, single global bf16 scale

Reference
BF16 PLE shard (weight-level comparison, serving/verify_fp8_ple.py)
Eval data
not stated (weight tensors, shard_0)
Tool
serving/verify_fp8_ple.py
  • Only fidelity figures are weight-level for the PLE FP8 conversion (shard_0 SNR, max relative error); no output-level quality numbers for the model [L55-56]
  • Expert quantization is cyankiwi's AWQ-INT4, not re-done by this publisher [L26]
  • PLE pinned-RAM footprint ~95 GiB -> ~48 GiB with CPU offload [L16]

Card snapshot c2aa55f204e1 · updated 2026-09-05

JRamirez-UAB/Qwen3.8-Flash-Next-EXL3-2.25bpw-48GB

EXL3 vLLM 0.31 + exllamav3 1.5.1 + own plugins, or TabbyAPI vs BF16

exllamav3 1.5.1 sc_optimize to 2.25 bpw (-mink 2: 83% experts 2-bit, 17% 3-bit), head 6, vision 6, MTP 4, 4-bit n-gram table; calibration 500 rows generated by NVIDIA's NVFP4 build

Reference
BF16 model
Eval data
internal sysadmin/agent evaluation set (unpublished) and turboderp's public qbench_prompts_gen.json
Tool
exllamav3 qbench
  • Internal KLD set is unpublished; public figure uses turboderp's trace [L91]
  • 37.68 GiB is the EXL3 weights without the n-gram table; download 66.7 GB incl. 26.2 GB EXL3 n-gram table; vLLM path uses NVIDIA's FP8 PLE table instead [L66-71]
  • Calibration data generated by NVIDIA's NVFP4 build of the model [L300-301]
  • Agent/tool tasks are in-house sysadmin tasks; card states ±2 of 30 run-to-run variation [L92]
  • Card's 48 GB memory figures come from the publisher's Swift build, not this checkpoint [L332]

Card snapshot 9eb541a11904 · updated 2026-10-16

Jundot/Qwen3.8-Flash-Next-oQ4e-mtp

MLX oMLX vs BF16

oQ4e: sensitivity-based bit allocation + imatrix (1024x512) + weighted LSQ scale refit

Reference
BF16
Eval data
oMLX benchmark corpora, 16 windows x 2,048 tokens per corpus
Context
2048
Tool
oMLX
  • Per-corpus KL only (code, novel); no single mean figure

Card snapshot c4ddfb54a967 · updated 2026-10-12

Kashapps/Qwen3.8-Flash-Next-Uncensored-ROCmFP4-FAST-GGUF

GGUF (ROCmFP4, fork-only) kingjones30/ROCmFPX fork (qwen4exp + ROCmFP4) modified weights other or no reference

Q4_0_ROCMFP4_FAST recipe from orcarouter/Qwen3.8-Flash-Next-Uncensored BF16 (abliterated); all ROCmFP4 except Q6_K output head

Reference
publisher's aligned (non-abliterated) build of the same tier, not BF16
Eval data
own prompt splits: 24 harmful, 12 harmless, 8 quality; prompts and completions not published
Tool
same harness, greedy
  • Weights are abliterated (orcarouter), not a faithful quant of Qwen's model [L65-67]
  • Quality evidence is an 8-item set with prompts not published [L150, L153]
  • Comparison is abliterated quant vs aligned quant of same tier, not vs BF16 [L144]
  • Card's patch download URL and field-report links point to kingjones777 repositories [L53, L103]; card text and measurement tables match kingjones777/Qwen3.8-Flash-Next-Uncensored-ROCmFP4-STRIX_LEAN-GGUF
  • Card's numbers match kingjones777/Qwen3.8-Flash-Next-Uncensored-ROCmFP4-FAST-GGUF's card (21 of 21 claim lines identical, detected automatically); treated as copied, not as an independent measurement

Card snapshot 1aa1c0999a6a · updated 2026-09-14

khuasar/Qwen3.8-Flash-Next-REAP-288-Q4g32-Q8-MTPLX

MLX MTPLX source checkout with local changes (not a released app) modified weights other or no reference

REAP-288 expert pruning (sh0wie's published selection: 288 of 512 routed experts per layer kept) from BF16; routed experts Q4 g32, sensitive matrices Q8 g64, n-gram table Q4 g32 on SSD; original 512-expert MTP head kept

Reference
none (no unpruned or BF16 comparison)
Eval data
eight-task in-house suite (invoice arithmetic, scheduling, Python, SQL, etc.), one attempt per task/mode
Context
65536
Tool
MTPLX (native)
  • Pruned: 224 of 512 routed experts per layer removed (43.75%) — not a faithful quant [L31, L78]
  • Only quality evidence is an 8-task single-run suite with no reference model; publisher calls it preliminary and claims no full-weight equivalence [L129-144]
  • Card says upload in progress and not to download until marked complete [L24, L307]
  • Expert selection calibrated on agentic-coding traffic only [L90, L291]
  • Disk 76.61 GiB includes the 29.802 GiB SSD n-gram table; main-model payload 44.390 GiB; MLX allocation 45.980 GiB after load [L54-59]

Card snapshot 05858670f3f9 · updated 2026-09-16

kingjones777/Qwen3.8-Flash-Next-Uncensored-ROCmFP4-FAST-GGUF

GGUF (ROCmFP4) kingjones30/ROCmFPX fork (qwen4exp + ROCmFP4) with bundled patches modified weights other or no reference

Q4_0_ROCMFP4_FAST: attention, experts, token embeddings and PLE all ROCmFP4 (TYPE_101), output head Q6_K; quantized from orcarouter's uncensored BF16

Reference
publisher's own aligned build of the same tier
Eval data
harmful (24), harmless (12), quality (8) prompts; prompts and completions not published
Tool
same harness, greedy (not named)
  • Quality evidence is an 8-prompt quality split (6/8, same two failures as the aligned build); prompts not published [L207-216]
  • Weights modified: abliterated by orcarouter; this repo only converts and quantizes [L96-97, L123-125]
  • Comparison is against the publisher's aligned ROCmFP4 build, not BF16 [L207]
  • GPU memory 63.3 GiB resident at full offload [L181]

Card snapshot ce1bddb037c4 · updated 2026-09-17

kingjones777/Qwen3.8-Flash-Next-Uncensored-ROCmFP4-FAST-imatrix-GGUF

GGUF (ROCmFP4) kingjones30/ROCmFPX fork (or charlie12345/ROCmFPX d3ca537 + bundled patches) modified weights other or no reference

Q4_0_ROCMFP4_FAST (TYPE_101, 4.251 bpw) on all tensors incl. PLE, Q6_K lm head; imatrix from bartowski calibration_datav3 computed on the 4-bit model; source orcarouter/Qwen3.8-Flash-Next-Uncensored BF16

Reference
none (no BF16 or base-model PPL; compared to own non-imatrix sibling)
Eval data
WikiText-2 raw test (wiki.test.raw)
Context
512
Tool
llama-perplexity (ROCmFPX)
  • Abliterated/uncensored source weights (orcarouter) [L5, L50, L149-151]
  • No BF16 reference: delta is vs the publisher's non-imatrix sibling only [L55-58]
  • imatrix computed on the 4-bit model, not BF16 (card states this) [L66-67]
  • Size 87.9 GiB is the model GGUF; shards listed as 44.70 + 44.67 + 5.06 GB, plus 0.91 GB mmproj [L60, L199-202]

Card snapshot 9abf014a574f · updated 2026-09-17

kingjones777/Qwen3.8-Flash-Next-Uncensored-ROCmFP4-STRIX_LEAN-GGUF

GGUF (ROCmFP4, fork-only) kingjones30/ROCmFPX fork (qwen4exp + ROCmFP4) modified weights other or no reference

Q4_0_ROCMFP4_STRIX_LEAN recipe from orcarouter/Qwen3.8-Flash-Next-Uncensored BF16 (abliterated); PLE Q5_1, token_embd Q5_K, output Q6_K

Reference
publisher's aligned (non-abliterated) build of the same tier, not BF16
Eval data
own prompt splits: 24 harmful, 12 harmless, 8 quality; prompts and completions not published
Tool
same harness, greedy
  • Weights are abliterated (orcarouter), not a faithful quant of Qwen's model [L123-125]
  • Quality evidence is an 8-item set with prompts not published [L212, L215]
  • Refusal/quality table and long-context table are identical to the card of Kashapps/Qwen3.8-Flash-Next-Uncensored-ROCmFP4-FAST-GGUF (a different tier) [L132-136 there, L194-198 here]
  • MTP speed figures were measured on other builds (FAST imatrix / STRIX_LEAN-imatrix weights) [L56-60, L76-78]

Card snapshot a4d9d7bfcbac · updated 2026-09-17

kingjones777/Qwen3.8-Flash-Next-Uncensored-ROCmFP4-STRIX_LEAN-imatrix-GGUF

GGUF kingjones30/ROCmFPX fork (qwen4exp + ROCmFP4 types, patches required) modified weights other or no reference

ROCmFP4 STRIX_LEAN mix (experts/shared TYPE_101, attention half TYPE_100, PLE Q5_1, token_embd Q5_K, output Q6_K) from orcarouter's abliterated BF16; imatrix on bartowski calibration_datav3, computed on the 4-bit model

Reference
none (no BF16 PPL; compared only to the publisher's FAST builds)
Eval data
held-out WikiText-2 raw (wiki.test.raw)
Context
512
Tool
llama-perplexity (fork)
  • Weights modified: quantized from orcarouter's abliterated (uncensored) BF16, not Qwen's weights [L5, L49, L144]
  • No BF16 reference PPL; only comparisons among the publisher's own tiers [L52-58]
  • imatrix computed on the 4-bit model, not BF16 [L72-73]
  • No same-recipe imatrix on/off delta for this tier; card cites the FAST tier's −5.9% instead [L64-67]
  • Chunk count not stated [L52, L210]

Card snapshot 503e614552a8 · updated 2026-09-17

klee100/Qwen3.8-Flash-Next-AutoRound-3bpw-MTP

AutoRound (GPTQ-format) patched vLLM a5a30471 (PLE SSD patch) vs BF16

Mixed precision: 2/3-bit routed experts (~2.998 effective bpw), 8-bit main projections, BF16 sensitive paths, 4-bit MTP experts

Reference
original weights (precision not stated)
Eval data
not stated
Tool
not stated
  • Single KLD figure with no dataset, context, token count or tool stated [L30]
  • 47.3 GiB is loaded VRAM; the 95.37 GiB BF16 PLE table stays on SSD [L39-40]
  • 2.998 bpw refers to routed experts only [L25]

Card snapshot 600715b31bed · updated 2026-09-24

kstoyanov99/qwen38-flash-next-nvfp4

NVFP4 SGLang (qwen4_exp) / vLLM with PLE disk-offload patch on DGX Spark vs BF16

RadixArk NVFP4 candidate (ModelOpt v0.46.0, routed experts W4A4, calib 128 cnn_dailymail @512) re-published with re-sharded PLE and mmap table parts

Reference
BF16 runs on an earlier checkpoint revision
Eval data
GSM8K full 1319 (t0.6, top-p 0.95, max 8192); AIME26 30 problems x 8 (t1.0, max 130k)
Tool
not stated
  • Re-publication of RadixArk's candidate weights; card text is RadixArk's [L19-26]
  • Table headed 'attributed to this exact checkpoint' but AIME26 was measured on a previous revision differing in PLE tables [L134, L145-146]
  • BF16 baselines from an earlier checkpoint revision [L141-143]
  • Card notes long agentic generations run longer than BF16 [L148-149]
  • Card's numbers match RadixArk/Qwen3.8-Flash-Next-NVFP4's card (11 of 11 claim lines identical, detected automatically); treated as copied, not as an independent measurement

Card snapshot ee25daef4ea1 · updated 2026-08-28

Kymidnight/Qwen3.8-Flash-Next-ROCmFP4-FAST-GGUF

GGUF LaurentZuijdwijk/llama.cpp fork (vulkan/qwen4exp-rocmfpx) vs BF16

ROCmFP4-FAST on experts/attention/GDN/hc, n-gram table Q3_0_ROCMFPX, token_embd/output Q6_K; no imatrix

Reference
unquantized PPL as reported in llama.cpp PR 27742 (not own run)
Eval data
wikitext-2 raw, 145 chunks
Context
2048
Tool
llama.cpp fork (perplexity)
  • Baseline PPL borrowed from llama.cpp PR 27742, not measured alongside [L75]
  • Publisher states the gap is wider than a good 4-bit quant should have; no importance matrix [L25-29, L78]
  • ~85 GiB VRAM when run [L53]

Card snapshot 6dd650011b94 · updated 2026-09-19

lackonendes/PAW-125B-FLASH-NEXT-X3.1

GGUF (TRUSS-only) TRUSS only (stock llama.cpp cannot read it) vs BF16

PAW X3.1 trellis codec (EXL3/QTIP-style, exllamav3 v1.5.1 encoder): 12,000 most-used routed experts 3.5 bit, other 12,576 at 2.5 bit (~3.0 bit avg); everything else Q8_0; n-gram table int8

Reference
BF16 base model run side by side; only routed experts taken from the quantized file, everything else BF16
Eval data
64 chunks x 2,048 tokens code/chat + 64 chunks WikiText-2, second half of each chunk scored (65,472 tokens per corpus); benchmarks GPQA-Diamond, LiveCodeBench v6, IFBench vs Qwen card BF16
Context
2048
Tool
not stated
  • KL isolates expert quantization: everything except routed experts kept BF16, so it does not measure the shipped file's Q8_0 dense weights or int8 n-gram table [L118-119]
  • GPQA-Diamond 92.4 is a union over up to four attempts per question, not a single-pass score, compared against Qwen's single figure [L137]
  • LiveCodeBench v6 run partial/in progress (131/175) [L134, L141]
  • IFBench ladder stopped after P5, P6 items counted as wrong [L140-141]
  • Limitations says 'Benchmarks are not in yet' while the Quality section lists benchmark results [L164, L131-135]
  • BF16 baselines are copied from Qwen's card, not run in the same harness [L131]
  • Sizes in KL table exclude the n-gram table; total download 95.8 GiB [L18] vs 96.8 GiB 'both files' [L63]
  • Plain-English KL uses WikiText-2

Card snapshot 257599e691ce · updated 2026-10-14

lee-chang-93/Qwen3.8-Flash-Next-NVFP4-REAP-k448

NVFP4 (ModelOpt) vLLM >= 0.30.0 / TRT-LLM / ModelOpt NVFP4 paths (num_experts=448) modified weights other or no reference

REAP expert pruning (keep 448 of 512 routed experts per layer) of nvidia/Qwen3.8-Flash-Next-NVFP4; kept weights bit-identical; calibration 16.49M tokens of the publisher's production agentic traffic

Reference
unpruned expert set (REAP saliency mass only; no output-level comparison)
Eval data
16.49M tokens of production traffic (681 LLM calls, 626 unique images)
Tool
REAP reference-semantics tooling
  • Weights modified: pruned (64 of 512 routed experts removed per layer) [L18-19]
  • Only fidelity figure is retained REAP saliency mass (92.65%; mean 7.35% removed per layer, worst 9.11%), a calibration proxy — no output-level quality numbers [L29, L47]
  • Card recommends QA diff vs unpruned baseline before production use [L64-65]
  • Pruned at 48% calibration coverage (16.49M of 34.2M tokens) [L62]

Card snapshot 8d565c90add2 · updated 2026-09-23

leoncca/Qwen3.8-Flash-Next-AWQ-g32

AWQ private V100-optimized vLLM build (TP4); generic AWQ support insufficient vs BF16

AWQ-format W4A16 g32 asymmetric, routed experts requantized from BF16 with damped GPTQ on captured activations (public80-v1mix20, 169,614 rows/layer); s1/s2 re-searched; PLE uses official FP8 weights; QSA FP8 KV scales preserved

Reference
BF16 official text model, transformers CPU full-vocab forward expanded to FP32
Eval data
fixed stratified 66-record text subset, 39,101 scored tokens
Tool
own collection via served vLLM + transformers CPU reference
  • KL is a top-20-plus-tail approximation (a lower bound); card's conservative full-KL bounds span [0.020465, 0.317644] [L147-156]
  • Comparison includes runtime arithmetic and KV effects (CPU FP32 reference vs GPU FP16 serving), not weight quantization alone [L126-128, L161-162]
  • Small eval: 66 records / 39,101 tokens [L123, L159]
  • Earlier GSM8K/HumanEval/MBPP/IFEval results describe previous weights and are not claimed for this release [L174-177]
  • PLE table is official FP8, not BF16 [L25-26]

Card snapshot ff2fb53a7e33 · updated 2026-10-17

lesj0610/Qwen3.8-Flash-Next-NVFP4

NVFP4 not verified (ModelOpt HF export) vs BF16

NVIDIA ModelOpt general NVFP4 experts-only recipe (constant input amax 2688, FP8 KV cast) + 4 disabled quantizers (PLE, conv1d, hyper-connections, QSA indexer); everything but routed experts BF16

Reference
original BF16 checkpoint weights (layer-0 experts 0/7/255/511)
Eval data
n/a (weight-space reconstruction, no inference)
Tool
own unpack/rescale comparison
  • Only weight-reconstruction fidelity (cosine/relative error on 4 sampled experts of layer 0); card states no inference was run and task quality is unmeasured [L163-170, L210-214]
  • Card reproduces Qwen's base-model tables below L245 [L245, L315]

Card snapshot e8045c65a50a · updated 2026-08-28

lhoestq/Qwen3.8-Flash-Next-NVFP4-Michel

NVFP4 vLLM (Blackwell B200/B300) other or no reference

Card is NVIDIA's NVFP4 card (ModelOpt v0.46.0: routed experts W4A4 NVFP4, attention/shared BF16, MTP experts FP8 128x128, PLE per-tensor FP8) with 'chat template modified for my personal use'

Reference
Qwen/Qwen3.8-Flash-Next-FP8 (not BF16)
Eval data
GPQA Diamond, HLE, τ²-Bench Telecom, MMMU Pro, SciCode, AA-LCR, IFBench, Omniscience, Terminal-Bench 2.1
Tool
not stated
  • Card copies nvidia/Qwen3.8-Flash-Next-NVFP4's card and evaluation table; numbers are NVIDIA's, not measured on this repo's files, which have a modified chat template [L24, L41]
  • Baseline is the official FP8 release, not BF16 [L115]
  • Sampled at temperature 1.0 with no run count or error bars [L115]
  • No absolute size, only '~2.7x smaller than BF16 (about 63% reduction)' [L188]
  • Usage commands point to nvidia/Qwen3.8-Flash-Next-NVFP4 [L199, L210]
  • Card's numbers match nvidia/Qwen3.8-Flash-Next-NVFP4's card (6 of 6 claim lines identical, detected automatically); treated as copied, not as an independent measurement

Card snapshot 81d3bdde1951 · updated 2026-09-22

Litwein/Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX

MLX oMLX modified weights vs BF16

REAP expert pruning 512->320 + oQ3e (3-bit g64 routed experts from oQ8e codes, M4Q init; 5/6/8-bit trunk; 4-bit g32 PLE) + end-to-end KL-DWQ v7->v9 against bf16 teacher top-20 logprobs; bf16 vision; REAP-matched MTP

Reference
bf16 Qwen3.8-Flash-Next teacher (top-20 logprobs incl. reasoning)
Eval data
41-record held-out table (same records every round); independent gate 17 held-out caches, 456,602 tokens; 120 held-out real-image vision records
Tool
custom end-to-end KL-DWQ eval (CE-gap, KL20)
  • Pruned: REAP keeps 320 of 512 routed experts per layer; card says pruning is lossy [L57-58, L64, L195]
  • Figures are teacher-fidelity (CE-gap, KL20 over top-20 logprobs) on held-out calibration distributions; card states they are not task accuracy and no benchmark score is claimed [L148-149]
  • Held-out sets are drawn from the same calibration distributions used for DWQ [L130, L148]
  • Prior-round values for gate/vision are labelled only as before->after; attributed here to v8 per the v8->v9 table context [L141-144]
  • bpw 4.24 overall, 3.76 excluding n-gram table; ≈36 GiB resident on 48 GB Mac [L106, L108]

Card snapshot 9a167839c453 · updated 2026-10-07

Litwein/Qwen3.8-Flash-Next-REAP320-oQ3e-fp16-DWQ-MTP-Vision-MLX

MLX (oMLX) oMLX modified weights vs BF16

REAP expert pruning 512->320 + oQ3e (3-bit g64 routed experts, M4Q init, 5/6/8-bit trunk, 4-bit g32 n-gram) + end-to-end KL-DWQ (v7->v9) against bf16 teacher top-20 logprobs; 16-bit tensors recast BF16->FP16, vision FP32

Reference
bf16 Qwen3.8-Flash-Next teacher, top-20 logprobs incl. reasoning tokens
Eval data
held-out calibration-distribution records: 41-record table (same records every round); independent gate 17 held-out caches, 456,602 tokens; real-image vision 120 held-out records
Tool
custom end-to-end KL-DWQ tooling for qwen4_exp (not named further)
  • Card states this FP16 repo shares weights, quantization, DWQ training and 'same evaluation numbers' with its BF16 sibling; teacher-gap tables describe both repos [L62-64]
  • Metrics are fidelity to the bf16 teacher (CE-gap | KL20 over top-20 logprobs) on held-out calibration distributions, not task accuracy; no task or vision-benchmark score claimed [L203-204]
  • Pruned: 192 of 512 experts per layer removed (REAP-320); card calls pruning lossy [L114-117, L255-256]
  • FP16 recast checked in value domain: 99.99973% of 5,787,423,569 FP16-typed values exact, 0 overflow, 148 flushed to zero [L87-94]
  • ~20% M1/M2 prefill gain is quoted from oMLX documentation; card says the speedup itself was not measured on this pack [L66, L103-106]
  • 33.6% recovery is of a 0.047-nat gap to an unpruned 512-expert 3-bit reference [L200-201]

Card snapshot 46a83dc51b5b · updated 2026-10-07

Litwein/Qwen3.8-Flash-Next-REAP320-oQ3e-fp16-DWQ-MTP-Vision-MTPLX

MLX (MTPLX layout) MTPLX >= 2.11.2 modified weights vs BF16

REAP expert pruning 512->320 + oQ3e (3-bit routed experts, mixed trunk) + end-to-end KL-DWQ; MTPLX repack with self-distilled 8-bit MTP sidecar; 16-bit tensors recast BF16->FP16, vision FP32

Reference
BF16 sibling repo's tensors (value-domain FP16 conversion check only)
Eval data
all 5,786,885,968 FP16-typed values in the pack
Tool
not stated
  • No task-quality claims for the quant: card states 'no task or vision-benchmark score is claimed yet' [L224]; teacher-gap tables are referenced on the MLX sibling card, not reproduced here [L117-118]
  • Recorded figures measure FP16 recast fidelity vs the BF16 sibling, not quantization error vs Qwen's BF16 [L84-95]
  • Pruned: REAP keeps 320 of 512 experts [L117]
  • Speed figures measured on the BF16 sibling pack on M4 Pro; no M1/M2 timing claimed [L164-166]

Card snapshot 0bb3ca69790f · updated 2026-09-18

local-inference-lab/Qwen3.8-Flash-Next-NVFP4

NVFP4 runtime supporting mixed NVFP4/MXFP8 ModelOpt layout incl. NVFP4 PLE (not named) modified weights other or no reference

Quantization-aware distillation (Quatrain trainer, 2,500 trunk + 1,500 joint updates) from BF16 teacher on 200M teacher-generated tokens; routed experts + PLE NVFP4, shared experts + attention MXFP8, rest BF16; static NVFP4 activations for routed experts

Reference
none (no BF16 row; compares to NVIDIA NVFP4)
Eval data
AA-LCR, GPQA Diamond, IFBench, Tool Eval
Tool
not stated (results linked to MiaAI-Lab GitHub PR #61)
  • Weights retrained (QAD distillation), not a post-training quant of the released weights [L22-27]
  • No BF16 or FP8 baseline row; NVIDIA comparison numbers are copied from NVIDIA's card (NVIDIA harness: temp 1.0, reasoning_effort=xhigh, FP8 baseline) rather than measured on the same harness [L109]
  • AA-LCR 79.4 has no source link (other three cells link to a GitHub PR) and exceeds NVIDIA's own reported FP8 baseline (71.9) by 7.5 points [L110]
  • Harness/settings for its own numbers not stated on the card
  • Tool Eval has no comparator (NVIDIA shows '—') [L109-110]
  • Disk size is not a runtime memory estimate (publisher note) [L100-101]

Card snapshot 6c01ac378d22 · updated 2026-10-06

LOKSUN-158/SamQuant-Qwen3.8-Flash-Next-Uncensored-MLX-Serve-Mix-C58

MLX (mlx-serve mixed) mlx-serve >= 26.9.1 (measured on 26.9.6 / MLX 0.32.2) modified weights vs BF16

Quantized from orcarouter's uncensored BF16 weights; imatrix calibration on publisher's own AI session logs + official general corpus; per-layer greedy bit allocation: routed experts 3-bit g64 x70 groups + 4-bit g64 x26 (layers 0-45), 4-bit g64 layers 46-47 + MTP experts; attention/GDN/shared expert 6-bit g64; lm_head 8-bit; n-gram table 4-bit g32 separate mmap file

Reference
BF16 original (top-1 agreement)
Eval data
600 held-out scoring positions, scoring prefixes 65~2,297 tokens; business acceptance 24 questions x 7 packs (tool orchestration / code / long docs, 8 each)
Tool
mlx-serve pipeline held-out window scoring
  • Weights modified: built from orcarouter's uncensored BF16; card says no further behavior changes by this publisher [L45, L109]
  • Top-1 agreement over only 600 positions with short prefixes (65~2,297 tokens); card says long-context quality not validated and top-1 can decouple from real task ability [L86, L108]
  • Comparison 4-bit community pack is not named [L87]
  • Card notes tool-orchestration acceptance 3/8 as a shared weakness of the packs tested [L90]
  • Calibration corpus is the publisher's own AI agent session logs [L47]
  • Size: ~92GB total; 59.83GB weights+vision plus 32,000,153,976-byte n-gram table; resident ~61GB after load [L62, L69]

Card snapshot a6692a8b6799 · updated 2026-10-11

Lottolabs/Qwen3.8-Flash-Next-Coder-P150

BFP4 (Tenstorrent) Strata tt backend (fnx) on Tenstorrent P150 only modified weights other or no reference

GPTQ re-quantization to Tenstorrent bfp4_b of the routed experts of ISTA-DASLab GSQ-RCO-Coder IQ1_M (256 of 512 experts kept); calib 128 x 2048 tokens (chat/code/instruct/docs)

Reference
the published ISTA-DASLab GGUF evaluated in float32 (not Qwen BF16)
Eval data
teacher-forced chat/code/doc records + GSM8K first 50; held-out, disjoint from calibration
Tool
Strata tt backend (P150)
  • Derived from ISTA-DASLab's pruned Coder GGUF (256 of 512 experts) — not a faithful quant [L23, L50-52]
  • Reference is the IQ1_M / IQ3_S GGUF in float32, not BF16 [L72-73, L81]
  • GSM8K on 50 items [L89]
  • Not a standalone model: no tokenizer, attention, embedding or LM-head weights; size is expert files only [L30-34]

Card snapshot 69e738b2c860 · updated 2026-10-13

lovedheart/Qwen3.8-Flash-Next-NVFP4-FP8

NVFP4 (+FP8 PLE) patched SGLang (lovedheart/qwen38-flash-next docker image) vs BF16

ModelOpt v0.46.0 NVFP4 W4A4 routed experts only; PLE tables from Qwen FP8 revision; rest BF16; calib 128 cnn_dailymail articles @512

Reference
BF16 runs on an earlier checkpoint revision
Eval data
GSM8K full 1319 (t0.6/top-p 0.95/max 8192); AIME26 30 problems x 8 (t1.0/max 130k)
Tool
not stated
  • BF16 baselines from an earlier checkpoint revision; card says treat as indicative [L228-230]
  • AIME26 figure measured on a previous revision of this checkpoint (PLE tables differ) [L232-233]
  • Card body describes itself as 'RadixArk Qwen3.8-Flash-Next-NVFP4' and its usage example loads RadixArk/Qwen3.8-Flash-Next-NVFP4; evaluation text and figures match that card [L110, L191, L220-236]
  • Card notes long agentic generations run longer than BF16 [L235-236]

Card snapshot a9786bf9d5bd · updated 2026-08-29

lovedheart/Qwen3.8-Flash-Next-NVFP4-FP8-Pruned-RTXPRO-6000

NVFP4+FP8 patched SGLang (lovedheart fork; stock SGLang emits garbled output) modified weights vs BF16

AIMER expert pruning 512 -> 448 per layer (calibration-free), NVFP4 routed experts + FP8 E4M3 128x128 blockwise weight-only attention/GDN projections

Reference
BF16 runs on an earlier checkpoint revision (as stated in the copied RadixArk evaluation section)
Eval data
GSM8K (1319), AIME26 (30x8)
Tool
not stated
  • Pruned (AIMER 512 -> 448 experts per layer): not a faithful quant [L21-32]
  • Body below L91 is RadixArk's NVFP4 card text: it names 'RadixArk Qwen3.8-Flash-Next-NVFP4', 512 routed experts, BF16 non-expert tensors, a 135 GB size and a RadixArk model path, which contradict this repo's pruned 448-expert, FP8-attention description [L65-69, L94, L120-122, L161-167, L175]
  • GSM8K/AIME26 numbers and footnotes are identical to RadixArk/Qwen3.8-Flash-Next-NVFP4's card; the card does not show they were measured on this pruned checkpoint [L193-203]
  • No size stated for this checkpoint (135 GB is in the copied RadixArk text) [L167]

Card snapshot 0e63a89e8ce4 · updated 2026-08-29

MagneticLab/Qwen3.8-Flash-Next-NVFP4

NVFP4 (ModelOpt) + NVFP4 PLE sidecar vLLM (vllm/vllm-openai:qwen38-flash-next image) with primitive-ai PLE overlay files reference not stated

ModelOpt NVFP4 on routed experts only; 51.2B PLE n-gram table served from a quantized (4-bit) sidecar mmapped from disk, quantized from the BF16 release

No evaluation method stated.

  • No output-quality claims for the quant (card read in full): only a weight-level PLE sidecar figure (worst_shard_rel_err 0.0896 vs 0.1493 for an unnamed 'reference build' quantized from FP8) [L231-233]; needle-in-haystack passes at 8k/25k/99k/185k [L178-179]; throughput tables [L167-199]
  • Full download 200 GB incl. a 99 GB duplicate PLE table never read at runtime; lean download 101 GB; checkpoint dir 186 GB -> 84 GB trimmed [L34-36, L76, L94]
  • Weights in memory 74.34 GiB (identical trimmed vs full) [L108]; elsewhere 'weights 75.9' at rest [L181]
  • vLLM overlay files are primitive-ai's work, bundled [L83-88]

Card snapshot e30ead75503c · updated 2026-09-09

Manav1011/Qwen3.8-Flash-Next-NVFP4

NVFP4 SGLang (qwen4_exp) vs BF16

NVFP4 MoE experts (ModelOpt v0.46.0); calib 128 cnn_dailymail articles @512; PLE from Qwen FP8 revision

Reference
BF16 runs on an earlier checkpoint revision
Eval data
GSM8K (1319), AIME26 (30x8)
Tool
not stated
  • Card is RadixArk's NVFP4 card text: it names 'RadixArk Qwen3.8-Flash-Next-NVFP4', a RadixArk release date and the --model-path RadixArk/Qwen3.8-Flash-Next-NVFP4 [L19, L40, L100]
  • Evaluation numbers and footnotes are identical to RadixArk/Qwen3.8-Flash-Next-NVFP4's card; the 'this exact checkpoint' attribution is RadixArk's wording [L116-128]
  • BF16 baselines from an earlier checkpoint revision; AIME26 measured on a previous revision [L123-128]
  • Card's numbers match RadixArk/Qwen3.8-Flash-Next-NVFP4's card (11 of 11 claim lines identical, detected automatically); treated as copied, not as an independent measurement

Card snapshot f86fa7242b47 · updated 2026-09-12

manjunathshiva/Qwen3.8-Flash-Next-tq4a-tq2e-g64

MLX (TurboQuant) turboquant-mlx-full >= 0.26.0 (stock mlx-lm cannot load) other or no reference

Data-free TurboQuant: randomized Hadamard rotation + Lloyd-Max codebooks; attention/lm_head 4-bit g64, routed experts 2-bit g64, n-gram/PLE + token embedding 2-bit affine g32, routers/indexer/gating full precision, vision tower bf16, MTP dropped

Reference
none (no BF16 run; PPL compared only to the publisher's own pre-rebuild weights)
Eval data
WikiText-2, 32 x 512-token chunks; 8-task battery (resident, temp 0.7, seeded)
Context
512
Tool
not stated
  • WikiText-2 PPL with no BF16 (or any other) reference; only compared to the earlier version of the same build [L224]
  • 8-task battery was measured on pre-rebuild weights, not the current files [L194-195]
  • Tiny eval: 32 x 512-token chunks PPL; 8-task battery [L192, L224]
  • Card itself says quality evidence is batteries on one machine, not a benchmark suite [L368]
  • Thinking-enabled runs close </think> within 5,000 tokens on only 7/15 seeds with current weights [L226, L363]
  • Size 52.00 GiB is the text model only, plus 0.84 GiB vision tower [L49]

Card snapshot 34811dcd31a3 · updated 2026-09-19

manuuuel/mica-qwen3.8-flash-next-q4

GGUF (Mica) Mica (Apple silicon, C) vs BF16

Importance-calibrated recipe: experts gate/up Q4_K, down MXFP4; dense/head/shared expert/n-gram key-value Q8_0; token embedding and n-gram table BF16 bit for bit; 252 calibration texts x 512 tokens

Reference
authors' unquantised weights run in the authors' code (single precision on CPU for the KL table)
Eval data
Mica's 100 quality texts x 512 tokens (25 WikiText-2 test, 25 CPython stdlib v3.14.7, 25 No Robots test in chat template, 25 Wikipedia in nine languages), 51,100 predicted positions; KL over the full 248,320-token vocabulary
Context
512
Tool
Mica make quality (GPU) and tools/quality/quality.py (authors' code)
  • Two measurement paths on the same 100 texts: Mica's GPU engine (nats/token, perplexity change, same top token) and the authors' code with this file's weights (adds KL mean and 99th percentile) [L93-116]
  • Code text shows slightly lower perplexity than the original (-0.11% / -0.08%) [L100, L113]
  • Calibration texts are stated to come from documents disjoint from the quality texts [L75-76]
  • Vision tower shipped as a separate BF16 GGUF (897,881,920 bytes); importance statistics file published [L28-29]

Card snapshot 354af7ad9d75 · updated 2026-10-06

matteiuspi/Qwen3.8-Flash-Next-W4A16-G128-300i

ModelSlim INT4 W4 g128 (Ascend 310P) OpenSensor vLLM + vLLM Ascend forks with custom 310P operators other or no reference

ModelSlim asymmetric per-group min/max RTN INT4 g128 on 73,728 routed-expert projections from BF16; rest FP16/FP32 (incl. PLE)

Reference
quality gate: none stated; GPQA: matched llama.cpp UD-IQ4_XS GGUF on one RTX 6000 Pro (not BF16)
Eval data
fixed 228-question zero-shot gate; GPQA Diamond 198 Q, AISBench zero-shot CoT prompts (dataset version b1ed2c), temp 0, seed 1024, top-k 20, top-p 1, 8,192-token output ceiling
Tool
native W4A8 service, 2x Atlas 300I Duo, TP4/EP4, MTP2
  • Card calls the GPQA run a serving-stack comparison, not an isolated quantization test; chat templates and formats differ [L182-183]
  • Benchmark extractor credits the last 'Answer: X' in unfinished reasoning: 26 Ascend / 20 RTX cases credited without a final answer [L193-197]
  • 81/198 Ascend responses hit the 8,192-token output limit [L190]
  • 228-question gate is a fixed sampled gate, not official full-dataset MMLU [L95-96, L426-427]
  • GPQA Ascend cases were resumed across server changes [L208-210]
  • No BF16 reference; quality gate not rerun after later runtime changes [L128-129, L259-260]

Card snapshot d32978ccdde1 · updated 2026-10-08

matteiuspi/Qwen3.8-Flash-Next-W8A8-DYNAMIC-300i

Ascend ModelSlim W8A8_DYNAMIC (ascend_v1) OpenSensor vLLM + vLLM Ascend forks (custom dev runtime), 4x Ascend 310P3 TP4 vs BF16

ModelSlim W8A8_DYNAMIC: routed-expert weights INT8 per-output-channel (FP32 scales/offsets), dynamic per-token INT8 activations in grouped matmul; embeddings, PLE, attention, GDN, shared experts, vision, lm_head FP16

Reference
BF16 reference (conversion-time NLL/PPL); RTX 6000 PRO llama.cpp UD-IQ4_XS run (GPQA)
Eval data
NLL/PPL: 28,245 next-token predictions on a paired text calibration set; GPQA Diamond first five items (AISBench 19018e9c, dataset b1ed2c, temp 0, seed 1024, 8,192-token output ceiling); seven-item semantic/tool smoke check
Tool
ModelSlim QDQ reference (NLL/PPL); AISBench (GPQA)
  • PPL/NLL is a W8A8 QDQ (simulated) reference on the calibration set, 28,245 predictions; card itself says the lower quantized PPL is not evidence of improvement and held-out evaluation is required [L315-325]
  • GPQA result for this checkpoint is only the first five (non-random) items; card calls it a pipeline smoke check, too small to estimate quality; full 198-item W8 replay pending [L331-348, L361]
  • Smoke check matched control at 6/7; outputs not bit-exact [L254-257]
  • 198-item GPQA figures are for the companion W4 checkpoint and the RTX UD-IQ4_XS reference, not this W8A8 checkpoint [L95-103, L360]
  • RTX IQ4_XS reference is given as 138/198 (69.70%) in the frozen-setting table but the W4 section says W4's 140/198 was 'equal to a matched RTX 6000 Pro IQ4_XS run' [L97-98, L360]
  • Export ~240.06 GB (223.57 GiB) because PLE table and non-expert state remain FP16 [L204-205, L431-432]

Card snapshot 27e5cdcbbcd1 · updated 2026-10-08

Maximilian228/Qwen3.8-Flash-Next-NVFP4-GPTQ-Strata

NVFP4 Strata NVFP4 fork only vs a Q8_0

NVFP4 experts by GPTQ + block-scale search from BF16; 57.5K calib tokens

Reference
Q8_0 (own build)
Eval data
5,153 tokens of code/text
Tool
Strata
  • KL 0.038 is stated in prose as 'table above', but no table on the card shows KL [L78]
  • Reference is the publisher's own Q8_0, not BF16
  • Small eval: 5,153 tokens
  • Runs only on the Strata fork; other publishers' files re-run inside Strata

Card snapshot 379ff0b8a39f · updated 2026-10-11

Maximilian228/Qwen3.8-Flash-Next-NVFP4-Q8-Mix-Strata

NVFP4+Q8_0 (Strata pack) Strata NVFP4 fork only other or no reference

Per-layer mix of the publisher's NVFP4 GPTQ experts and Q8_0 experts (byte copies), chosen by a KL model fitted to 9 measured packs under a 110 GiB RAM budget; dense GGUF Q8_0/BF16, Qwen FP8 PLE

Reference
publisher's own Q8_0 Strata build (Q8_0 experts and attention, BF16 router, Qwen FP8 PLE), not BF16
Eval data
5,153 tokens of code, docs, English and Ukrainian chats and agent transcripts; plus thinking text written by the GPTQ model
Tool
Strata NVFP4 0.1.42-nvfp4.1 (experts on CPU, route tail skip off)
  • Reference is the publisher's own Q8_0 build, not BF16 [L40, L51-55]
  • Small eval: 5,153 tokens [L51]
  • Thinking-text column uses text written by the GPTQ model, which the card notes favours GPTQ [L115-116]
  • Comparison table sizes are RAM in use (engine + server); disk download is 171.6 GB (159.8 GiB) [L59, L130]
  • Other publishers' files were re-run inside Strata with route tail skip off, not in their own runtimes [L52-53, L70-71]
  • Runs only on the Strata NVFP4 fork [L98-99]

Card snapshot 9ac833ee7936 · updated 2026-10-11

Maximilian228/Qwen3.8-Flash-Next-Q8_0-Strata

Strata pack (Q8_0 experts) Strata NVFP4 fork only vs a Q8_0

All routed experts Q8_0 (RTN from BF16); dense GGUF Q8_0/BF16, Qwen's FP8 PLE table, BF16 embedding, MTP head (shared with the NVFP4 GPTQ repo)

Reference
this model (Q8_0) is the reference for the comparison table; expert error vs BF16 cited from the GPTQ card
Eval data
5,153 tokens of code, docs, English and Ukrainian chats and agent transcripts
Tool
Strata (experts on CPU, route tail skip off)
  • This quant is itself the reference: its 100% top-1 is by definition; only its quality figure vs BF16 is the expert relative MSE 0.00006 (citing the GPTQ card) [L37, L77]
  • Comparison reference is Q8_0, not BF16 [L26, L48]
  • Small eval: 5,153 tokens [L48]
  • Sizes in the table are RAM in use, not disk; disk is 189.2 GB (176.2 GiB) [L35, L99]
  • GGUF models were run with route tail skip on by default; card says their answers are a little further from Q8_0 than the table shows [L67-68]
  • Other publishers' files re-run inside the Strata fork

Card snapshot d8fc7dfc62f3 · updated 2026-10-11

mbehr90/Qwen3.8-Flash-Next-fp8

FP8 (compressed-tensors, FP8_DYNAMIC) vLLM with Qwen4ExpForConditionalGeneration (out-of-tree plugin; not vLLM 0.26.0) vs BF16

FP8 per-output-channel weights, dynamic per-token activations, routed experts only (quantized shard by shard without building the model, no calibration); PLE, embeddings, attention/GDN, shared experts, gates, vision, MTP, lm_head bf16

Reference
bf16 source (own run)
Eval data
GSM8K, 200 questions, via chat endpoint with final number parsed from response
Context
8192
Tool
vLLM on 4x H100 80 GB (serving command max-model-len 8192); custom parsing, not lm-eval strict-match
  • Single benchmark, n=200; card states SE ~1.5 pp and all three statistically indistinguishable [L26-33]
  • GSM8K answers parsed from chat responses by the publisher, not lm-eval strict-match [L33-35]
  • Size 220.8 GiB vs 335.3 GiB bf16 source [L20]
  • NVFP4 row is the publisher's own sibling build (repo name not given on the card) [L30]

Card snapshot e70592b7a40d · updated 2026-08-27

mbehr90/Qwen3.8-Flash-Next-nvfp4

NVFP4A16 (compressed-tensors) vLLM with Qwen4ExpForConditionalGeneration plugin (Marlin on H100) vs BF16

NVFP4A16 weight-only (fp4 weights, fp8 group-16 scales, bf16 activations) on the 48x512 routed experts, quantized shard by shard without calibration

Reference
bf16 source, same table
Eval data
GSM8K, 200 questions, via chat endpoint with final number parsed
Tool
own chat-endpoint script (lm-eval strict-match not used); 4x H100
  • Single benchmark, n=200; card states standard error ~1.5 pp and all three statistically indistinguishable [L32-33]
  • Quantized model scores above the bf16 source (0.960 vs 0.950), within the stated error [L28-30]
  • No calibration (weight-only scheme) [L22]

Card snapshot d28eef6b7a1d · updated 2026-08-27

meldlabs/Qwen3.8-Flash-Next-Meld-Turbo-Quality-GGUF

GGUF (EXL3 expert tensors) Meld Turbo engine >= 0.1.1 only vs a Q8_0

Repack: routed experts EXL3 2.625 bpw copied from Sushi 2.6 bpw pack; attention/linear-attention/shared experts/embeddings/output Q8_0 from unsloth GGUF; n-gram table IQ4_NL from ISTA-DASLab GSQ-RCO

Reference
Q8_0 (PPL 2.108)
Eval data
wikitext, 16 x 512 tokens; GSM8K 200, HumanEval 164, MMLU 200
Context
512
Tool
not stated
  • Reference is Q8_0, not BF16 [L44]
  • Tiny PPL eval: 16 x 512 tokens of wikitext [L44]
  • Not an original quant: tensors copied from Sushi (EXL3), unsloth (Q8_0) and ISTA-DASLab (IQ4_NL) [L33-36]
  • Lower-quality 'fast' profile scores higher on GSM8K (98.0% vs 96.5%) [L45]

Card snapshot ecbcece0c263 · updated 2026-10-17

mensaprodigy/Qwen3.8-Flash-Next-REAP-288-oQ4e-BF16-MTP-PLE

MLX oMLX v0.7.0.dev4 (qwen4_exp) modified weights other or no reference

REAP-pruned 288 target experts per layer, oQ4e 4-bit g64 with 232 overrides, native 512-expert MTP retained

Reference
none (no unpruned or BF16 run)
Eval data
HumanEval 100-question sample of 164; LiveCodeBench 100 of 1,054; seed 42 stratification
Tool
oMLX v0.7.0.dev4 (batch 4, TurboQuant KV 4-bit, Lightning MTP, M4 Max)
  • Pruned: 224 of 512 experts per layer removed — not a faithful quant
  • Accuracy from 100-question samples; scores belong to a local 'matching source artifact', not stated to be these Hub files [L35, L44]
  • Repo name says BF16 but the config is mixed 4-bit [L29-31]
  • Cross rows are quoted from other publishers' cards / oMLX records with different harnesses; card says not a controlled ranking [L66, L79]
  • No size stated

Card snapshot 0ee18f705397 · updated 2026-09-21

mensaprodigy/Qwen3.8-Flash-Next-REAP-384-mlx-mtp

MLX (oMLX oQ4e) oMLX main (needs PR #3659 + MTP stacking fix) modified weights no quality claim

sh0wie REAP-384 BF16 target (512 -> 384 experts per layer) + original 512-expert native MTP, converted with oMLX oQ4e (4-bit g64 primary, 392 overrides; PLE 4-bit g32)

No evaluation method stated.

  • Pruned (REAP 512 -> 384 experts per layer): not a faithful quant [L19, L38]
  • No quality measurement of this checkpoint; the HumanEval table is sh0wie's reported lineage figures, explicitly 'not direct measurements of this release' [L127, L165]
  • Lineage figures are single-run HumanEval (164 problems) with no confidence intervals, per the card [L136]
  • Throughput numbers are service counters, not a controlled benchmark [L140-146]
  • No size stated on card

Card snapshot ae3a5046145a · updated 2026-09-15

mensaprodigy/Qwen3.8-Flash-Next-REAP-384-oQ4e-BF16-MTP-PLE

MLX (oMLX) oMLX v0.7.0.dev4 (qwen4_exp, asymmetric target/MTP expert counts) modified weights other or no reference

REAP expert pruning to 384 of 512 experts per layer (MTP predictor keeps 512); oQ4e mixed 4-bit affine g64 with 260 named 5/6/8-bit overrides

Reference
none
Eval data
HumanEval 100 of 164 problems; LiveCodeBench 100 of 1,054 questions (seed 42 stratified samples)
Tool
oMLX benchmarks
  • Pruned (REAP): 128 of 512 routed experts removed per layer — not a faithful quant [L16, L24]
  • Repo name says BF16 but card states mixed 4-bit with 5/6/8-bit overrides, not BF16 [L30, L32]
  • No reference (BF16 or unpruned) compared; 100-question samples only [L36, L92]
  • No size stated; only peak memory 58.7–61.6 GB in throughput runs [L68-70]

Card snapshot 57cccc3abd84 · updated 2026-09-21

mensaprodigy/Qwen3.8-Flash-Next-REAP-384-oQ5e-BF16-MTP-PLE

MLX oMLX v0.7.0.dev4 (qwen4_exp) modified weights other or no reference

REAP-pruned 384 target experts per layer, 5-bit g64 with 303 overrides, native 512-expert MTP retained

Reference
none (no unpruned or BF16 run)
Eval data
HumanEval 100 of 164; LiveCodeBench 100 of 1,054; seed 42 stratified sampling; Thinking No / Yes runs
Tool
oMLX v0.7.0.dev4 (batch 4, TurboQuant KV 4-bit, Lightning MTP, M4 Max)
  • Pruned: 128 of 512 experts per layer removed — not a faithful quant
  • All accuracy values are 100-question samples [L35, L123]
  • Repo name says BF16 but the config is mixed 5-bit [L29-31]
  • Cross rows are quoted from other cards / oMLX records with different harnesses and batch sizes; card gives no delta [L56-58]
  • No size stated

Card snapshot ff3283772f74 · updated 2026-09-21

Minachist/Qwen3.8-Flash-Next-INT4-Mixed-AutoRound

AutoRound (compressed-tensors) patched vLLM (vllm-patch/, nightly 0.29.1rc1.dev47) vs BF16

Mixed INT4/INT6/INT8 AutoRound: routed experts INT4 gs128, linear_attn/QSA/shared expert INT6 gs64, HC mixers INT8 gs64, lm_head/embeddings/PLE projections INT8 gs128; router, vision, PLE table, MTP BF16; new-main recalibrated with data generated by a higher-precision version of this model

Reference
BF16 (KLD); benchmarks vs official BF16 model-card figures
Eval data
KLD: one English trace, response positions scored (14,496 input + 45,588 output tokens); IFBench 300 prompts, GPQA Diamond 198 questions, LiveCodeBench v6 131 problems (2025-02-01 to 2025-04-06)
Tool
qbench (exllamav3); IFBench evaluation_lib; simple-evals prompt; lcb_runner template
  • Benchmark baselines copied from the official model card, not run in the same harness [L76]
  • LiveCodeBench v6 uses 131 problems up to 2025-04-06, whereas the official label is 25.02-25.05 — different problem set from the BF16 figure [L80]
  • Benchmarks are for main only, not the recommended new-main [L68]
  • Card gives ±1 SE of 2.1–2.3; all deltas vs BF16 are within 1 SE [L70-74]
  • Unsloth GGUF KLD points are only in a chart image, taken from turboderp's published chart, not measured by this publisher [L53, L62]
  • bpw 4.178 is body bits per weight; no total size stated (BF16 PLE table file 95.37 GiB) [L38, L55]

Card snapshot 9bea61dda1a9 · updated 2026-10-09

mk2878d/Qwen3.8-Flash-Next-GGUF

GGUF llama.cpp with qwen4exp support (PR #27742) vs BF16

Card text is AtomicChat's (quantized_by: AtomicChat): AD builds, imatrix on BF16 (4000 chunks @512, 4,967,044 tokens), n-gram table in its own shard; asymmetric high-bit band on blocks 0-3 and 40-47

Reference
BF16 model's own logits
Eval data
held-out neutral set, 87 chunks
Context
4096
Tool
not stated (KLD logs in AtomicChat metrics repo)
  • Card is a copy of AtomicChat's card (front matter quantized_by: AtomicChat; links and commands point to AtomicChat/Qwen3.8-Flash-Next-GGUF) [L8, L29, L290]
  • AD-4.27bpw row here says 54.5 GB in memory / 92.9 GB total, whereas the AtomicChat card version in this dataset lists 56.1 GB / 94.5 GB [L83]
  • Says 4.27bpw is '17.6 GB smaller' than 5.00bpw; 110.5 − 92.9 = 17.6 [L87]
  • Headline '63% KLD reduction at the same file size' compares the 84.9 GB build with the 92.9 GB build [L188-189, L82-83]
  • Card says other publishers' files were re-measured, but no cross rows appear on this card [L96-98]
  • 'Choosing a quant' table lists IQ2_M/IQ3_M/Q4_K_M/UD-Q4_K_XL/Q6_K/Q8_0 with no sizes ('still uploading') [L270-280]
  • Card gives 177B params in one place and 125B + 51B + 4B MTP in another [L43, L256]
  • Card's numbers match AtomicChat/Qwen3.8-Flash-Next-GGUF's card (46 of 51 claim lines identical, detected automatically); treated as copied, not as an independent measurement

Card snapshot 69652fb0c020 · updated 2026-09-27

mudler/Qwen3.8-Flash-Next-APEX-GGUF

GGUF llama.cpp (recent, qwen4exp support) vs BF16

APEX: 32-block types forced on per_layer_token_embd and ffn_down_exps; n-gram table IQ4_NL; layer-position and gate<up bit allocation on ffn_gate/up_exps; imatrix on BF16 at 4096 ctx (chat, code, reasoning, tool calling, agentic, Wikipedia)

Reference
BF16
Eval data
same corpus the importance matrix was built from (named only as diverse calibration data)
Context
4096
Tool
not stated (perplexity)
  • Eval text is the imatrix calibration corpus; publisher says this flatters both quants and a held-out set would be fair [L96]
  • Tiny eval: six 4096-token chunks (~24.6K tokens) [L86]
  • Compact tier has no quality numbers [L41, L88-92]
  • mem = VRAM; n-gram table (29.1 GB) stays in system RAM [L37-41]
  • Side ablation (n-gram table IQ4_NL vs Q8_0: KLD 0.0324, +0.71% PPL, 6% top-1) is on 100 chunks held-out wikitext-2, a different set from the PPL table [L110]

Card snapshot d0ab3614ca01 · updated 2026-09-11

munekazu/Qwen3.8-Flash-Next-exl3-CMP170HX-64GB

EXL3 ExLlamaV3 v1.5.1+ / TabbyAPI other or no reference

No new quantization: module-wise mix of turboderp's 4.05bpw_h6_ng6 and 3.05bpw_h5_ng5 EXL3 tensors (4-bit experts on layers 0-9 and 38-47, 3-bit on 10-37; experts avg 3.43 bpw; 6-bit attention/shared/lm_head; n-gram table K=6)

Reference
none
Eval data
HumanEval 164 (base and plus tests), GSM8K first 300 test problems
Tool
not stated
  • No reference model (BF16 or turboderp source branches) compared [L57-63]
  • Not an original quant: every tensor comes from turboderp/Qwen3.8-Flash-Next-exl3 [L67, L153]
  • GSM8K is the first 300 test problems only [L63]
  • bpw 3.43 is the experts' average, not whole model; no total size stated (n-gram table 39 GB) [L28, L81]

Card snapshot b429d2768ee6 · updated 2026-09-21

mxxm-t/Qwen3.8-Flash-Next-MTP-GGUF

GGUF (MTP draft head only) mxxm-t/mx-llama.cpp fork vs BF16

NextN/MTP draft head extracted from mtp.* tensors via convert_hf_to_gguf.py --mtp; Q8_0 and BF16

Reference
BF16 head file (draft-token identity)
Eval data
prompts tested (not listed), temperature 0
Tool
llama-server --spec-type draft-mtp
  • Not a full model: draft head only (~2.6B params plus embeddings) [L28-31]
  • Only fidelity claim: Q8_0 and BF16 heads give identical draft tokens at temperature 0 (181 drafted, 136 accepted) [L40-41]

Card snapshot c7c423f1d0e1 · updated 2026-08-28

myllmbox/Qwen3.8-Flash-Next-hibrid47

NVFP4 mixed (vLLM) vLLM with myllmbox PLE patch, 2x DGX Spark TP=2 other or no reference

Routed experts NVFP4 W4A4 from Inferact; GDN projections NVFP4 W4A16; attention/shared/lm_head bf16; PLE n-gram table NVFP4 (28.6 GiB)

Reference
none (comparison vs own hibrid46 with int3 table)
Eval data
32-scene 'boss-animals' render gauntlet, thinking on; scoring criteria not stated
Tool
vLLM, 2x DGX Spark GB10, MTP K=4, temperature 1.0
  • Custom render test; pass/partial/broken criteria not stated [L30]
  • Ladder score is qualitative ('about half/half') [L31]
  • No BF16 reference; no total checkpoint size stated (PLE table 28.6 GiB) [L43]
  • Routed-expert tensors derived from Inferact/Qwen3.8-Flash-Next-NVFP4 [L38, L102]

Card snapshot 7b83fa0d4e55 · updated 2026-09-06

myllmbox/Qwen3.8-Flash-Next-hibrid48

NVFP4 vLLM 0.29 + myllmbox patch (lm_head quant_config), 2x DGX Spark TP=2 other or no reference

hibrid47 body (Inferact NVFP4 routed experts, myllmbox NVFP4 W4A16 GDN projections, NVFP4 PLE table) with lm_head newly NVFP4 W4A16 group 16

Reference
none (A/B against hibrid47 / bf16 head pending)
Eval data
HumanEval complete (164); fixed 200-question subsets (seed 123123123) of GSM8K, IFEval, MMLU-Pro; thinking on, 32k-token budget, Qwen thinking sampling 0.6/0.95/20
Tool
lm-evaluation-harness against the running serve
  • No baseline: publisher states these are the model's numbers, not yet a measured delta against the bf16 head [L75-77]
  • 200-question subsets; 11 of 763 answers hit the 32k budget and count as wrong [L66, L75]
  • Size is approximate single-Spark load (~101 GiB); ~50 GiB per box when split over two [L96]
  • lm_head quantization error stated as mean |Δw| = 9.0 % of |w| [L40]

Card snapshot e316057ddc79 · updated 2026-09-13

Myric/Qwen3.8-Flash-Next-APEX-GGUF

GGUF llama.cpp mainline (stock b2276-22397c31a verified) other or no reference

MIDDLE mixed recipe: PLE table and ffn_down_exps Q4_0 (shape-forced), gate/up experts IQ4_XS (40) / IQ3_XXS (56), dense Q6_K/Q8_0; imatrix 4,000 chunks; MTP head dropped

Reference
none
Eval data
agentic coding suite frontier tier, 9 tasks, one run; 5 vision perception tasks with blind control
Context
65536
Tool
own harness, pytest grading
  • No reference comparison; card calls the 9/9 sweep 'a floor rather than a ranking', one run [L142-144]
  • Perplexity deliberately withheld because the wikitext copy used was not the canonical artifact [L148-154]
  • MTP head (2.607 B params) is absent from both files [L175-184]
  • mem = ~85.4 GiB resident for weights [L62]

Card snapshot daaeccd6c619 · updated 2026-09-09

nazeshinjite/Qwen3.8-Flash-Next-Q8-UDQ4K-DS4

GGUF (ds4) DwarfStar (ds4) Metal other or no reference

Repack of antirez Q4: all ffn_down_exps to Q8_0 (from unsloth UD-Q5_K_XL shards), gate/up Q8_0 in layers 0,1,2,45 (unsloth Q8_0 shards), other gate/up swapped to unsloth UD-Q4_K_XL imatrix Q4_K; BF16 n-gram table unchanged

Reference
Alibaba API official continuations (top-five logprobs), not BF16
Eval data
qwen38-flash-alibaba-100 (100 prompts, 5,696 target tokens, ctx 4096) and qwen38-flash-alibaba-long (12 continuations, 766 target tokens, ctx 32768)
Context
4096
Tool
ds4 gguf-tools/quality-testing/score_official; llama.cpp control scorer score_llama for Unsloth files
  • Reference is Alibaba API top-five logprobs, not BF16; card states a BF16 KLD is not obtainable for this file [L82, L121-123]
  • Unsloth files scored with a llama.cpp control scorer, this file with ds4 — different runtimes in one table [L103-110]
  • Headline claims 'Equivalent quality to Unsloth's UD-Q5_K_XL'; table shows Q5 0.0011 NLL better on the short set, CI spanning zero [L19, L108, L114]
  • Small long set: 12 continuations, 766 target tokens [L82]
  • Not an original quant: tensors taken from unsloth and antirez files [L69-75]
  • 188.57 GiB file includes 95.37 GiB BF16 n-gram table read from disk; 93.19 GiB resident [L25, L28]

Card snapshot 9374dd4e1f8d · updated 2026-09-17

nibor1896/Qwen3.8-Flash-Next-CNQ4.5-M

CNQ (NVFP4) crow-nest only (RTX 5090, sm_120) other or no reference

CNQ4.5-M: NVFP4 4.5 bpw RTN, calibration free, mse sub-block scales; BF16 for embeddings, lm_head, router, shared-expert gate, norms, 1-D tensors

Reference
f32 originals (oracle KLD, pre-fix engine); llama.cpp UD-Q2_K_XL for the ten-task gate
Eval data
oracle KLD on 298 and 607 rows; ten-task gate (greedy and six sampled seeds)
Tool
crow-nest engine (commit 09cb00a for KLD)
  • Oracle KLD measured on the engine before the 2026-09-23 PLE-row fix; card says container KLD against BF16 is not measured [L418, L562]
  • Tiny eval: KLD on 298/607 rows; ten-task gate [L414, L418]
  • Ten-task runs predate the PLE fix; publisher notes greedy crow-nest is one task below the llama.cpp reference [L417]
  • Comparisons are against a different, smaller quant (UD-Q2_K_XL, 2.4 bpw), not BF16 [L383]
  • Downstream benchmarks not measured [L560]

Card snapshot ce0ddccda55b · updated 2026-10-07

nirecom/Qwen3.8-Flash-Next-mixed-3_8bit

MLX mlx_vlm / LM Studio MLX runtime other or no reference

Mixed precision adapted from pipenetwork's layout: routers/MoE gates and QSA indexer unquantized; PLE table 4-bit g32; routed experts 3-bit g64; everything else 8-bit g64

Reference
none (absolute IFEval scores; external comparison to unsloth UD-Q3_K_XL)
Eval data
IFEval, full 541 prompts, 0-shot
Tool
lm-evaluation-harness 0.4.13 local-chat-completions, greedy; max_gen_toks 1280 (thinking off) / 8192 (thinking on, --thinking-budget 4096); M5 Max 128 GB
  • IFEval only (instruction-following); card states knowledge, reasoning, coding, multilingual, safety and vision not measured [L71-75, L140-145]
  • No BF16 baseline; comparison point is unsloth UD-Q3_K_XL run through the same harness [L84-85]
  • Thinking-off: unsloth UD-Q3_K_XL slightly ahead on all four metrics; thinking-on: this build slightly ahead [L89-92, L96-97]
  • Thinking-on rows used different reasoning caps per runtime (MLX --thinking-budget 4096, llama.cpp --reasoning-budget 4096; GGUF had 2 empty answers of 541) [L106-108]
  • 85 GB, 4.094 bpw [L27-28]

Card snapshot a158876aa37c · updated 2026-09-26

nitinpanj/qwen38-flash-next-v3

GGUF npanj/llama.cpp fork only (--moe-stream expert streaming) other or no reference

Splice: bartowski Q4_0 + Q8_0 output.weight + unsloth UD-IQ4_XS for attn, hc, token_embd, ssm_out, shexp tensor groups

Reference
the unspliced checkpoint (not BF16)
Eval data
not stated (paired 40-chunk perplexity)
Tool
not stated (perplexity)
  • No BF16 reference: PPL is compared only against the unspliced checkpoint [L159-163]
  • PPL dataset and context not stated [L159]
  • Not an original quant: tensors are bartowski's Q4_0 and unsloth's UD-IQ4_XS [L156-157, L189]
  • Does not run on stock llama.cpp [L27-28]

Card snapshot e2982050848c · updated 2026-09-19

nota-ai/Qwen3.8-Flash-Next-Nota-NVFP4

NVFP4 vLLM qwen38-flash-next image (stock), Blackwell; --moe-backend flashinfer_cutlass or marlin vs BF16

NVFP4 W4A4 g16 compressed-tensors on routed experts only; calib 512 x 4,096-token conversations (agentic/SWE/code, 29.30% Korean)

Reference
BF16 (own run)
Eval data
GPQA-Diamond, HumanEval, AIME 2026, HRM8K (KSM subset, 1,428 of 8,011), BFCL v4 (13 single-turn categories)
Context
131072
Tool
not stated
  • NVFP4 scores above BF16 on 4 of 5 benchmarks; publisher states each difference is within sampling error and means absence of degradation [L92-96]
  • Single greedy pass, no repeats or error bars [L75-78]
  • HRM8K and BFCL restricted to subsets [L80-90]
  • Size is the computed checkpoint total (335.28 -> 173.56 GiB) [L31, L134]

Card snapshot 53f1bea6324e · updated 2026-09-02

nvidia/Qwen3.8-Flash-Next-FP4-Local

NVFP4 (+MXFP8) vLLM community containers only (eugr/spark-vllm-b12x, local-inference-lab) modified weights vs BF16

Quantization-aware distillation (QAD) with ModelOpt v0.47.0: routed experts NVFP4, attention MXFP8; 2,500 trunk + 1,500 joint-refinement updates also training n-gram tables; 500M tokens x 3 epochs

Reference
BF16 (Qwen/Qwen3.8-Flash-Next)
Eval data
GPQA Diamond (198 questions), ToolEvalBench
Tool
not stated
  • Weights trained via quantization-aware distillation (gradient updates incl. n-gram tables), not a post-training quant [L23, L123-127]
  • No sampling settings, run counts or error bars for the two benchmarks [L106-120]

Card snapshot b9ce18a6887e · updated 2026-10-15

nvidia/Qwen3.8-Flash-Next-NVFP4

NVFP4 vLLM (Blackwell B200/B300) other or no reference

ModelOpt v0.46.0: routed experts W4A4 NVFP4 with MSE-calibrated weight scales; attention/shared experts BF16; MTP experts FP8 128x128; PLE per-tensor FP8 (copied from Qwen FP8); calib cnn_dailymail + Nemotron-Post-Training-Dataset-v2

Reference
Qwen/Qwen3.8-Flash-Next-FP8 (not BF16)
Eval data
GPQA Diamond, HLE, tau2-Bench Telecom, MMMU Pro, SciCode, AA-LCR, IFBench, Omniscience, Terminal-Bench 2.1
Tool
not stated
  • Baseline is the official FP8 release, not BF16 [L113]
  • Sampled at temperature 1.0 with no run count or error bars; NVFP4 scores above FP8 on 5 of 9 benchmarks, and differences this small cannot be separated from sampling noise without repeat runs [L67-108, L113]
  • Card gives no absolute size, only '~2.7x smaller than BF16 (about 63% reduction)' [L186]; size taken from HF listing
  • Card documents an 8-GPU TP8 deployment [L197-198]

Card snapshot fc694b54fb01 · updated 2026-09-05

octolix/Qwen3.8-Flash-Next-Octojet-NVFP4

NVFP4 experts + MLX affine 4-bit (Octojet-only) Octojet 0.1.x (DGX Spark / GB10) other or no reference

Byte-for-byte splice: routed experts NVFP4 from RadixArk/Qwen3.8-Flash-Next-NVFP4 @ 7b719225; shared experts bf16 (packed to NVFP4 at load); attention/DeltaNet/routers/embeddings/n-gram/head/MTP MLX affine 4-bit g32 from Vontra/Qwen3.8-Flash-Next-MLX-4bit-MTP @ 2b170fa6; no requantization

Reference
none (absolute scores; compared against TensorFold v0.6.0 + local-inference-lab NVFP4)
Eval data
GSM8K 250 problems; HumanEval pass@1 164 programs
Tool
not stated on card (docs/benchmarks.md linked)
  • Accuracy measured on Octojet's production build whose shared experts came from an FP8-derived copy, not the bf16 shared experts shipped in this repo [L115-117]
  • Card notes the 3-program HumanEval gap is within noise at 164 [L111-112]
  • Not an original quant: tensors copied byte-for-byte from RadixArk NVFP4 and Vontra MLX 4-bit [L126-131]
  • No disk size stated; 86.2 GiB is a startup memory estimate [L81]
  • Most of the card is speed comparisons against TensorFold on other checkpoints [L48-97]

Card snapshot a713d8c02110 · updated 2026-10-15

OsaurusAI/Allosaurus-v0.1-125B-A6B-JANGH2

MLX (JANGH) Osaurus >=0.25.18 vs BF16

JANGH 2-bit routed experts (codebook + per-row scales + Hadamard-32, GPTQ per expert covariance); resident weights affine 4/8-bit with imatrix + AWQ; n-gram table affine 8-bit on SSD; 607k-token calibration

Reference
full bf16 model run layer by layer from disk
Eval data
four held-out sets (general 34 seqs, reasoning traces 120 seqs, tool conversations 96 seqs, office/finance/math 400 seqs), 514,498 teacher-forced positions
Tool
own scorer (Osaurus)
Unit
top-128 renormalized KL
  • KL is top-128 renormalized, not full-vocabulary [L63]
  • Reference row is a 'ceiling' (bf16 with injected noise), not exact bf16 parity; card frames cost as distance to the ceiling [L65-67]
  • Size is resident weights; 54 GiB n-gram table read from SSD [L37, L151]
  • Long-documents domain: previous release is closer to bf16 than this one [L100-102]
  • Reasoning-effort comparison uses the publisher's 6-bit release as reference, not bf16 [L126]

Card snapshot 01d629ec56eb · updated 2026-10-04

otheru/Qwen3.8-Flash-Next-Strix-Halo-GGUF

GGUF + halogen .hgn halogen-flash-server 0.12.1 (gfx1151 only); does not load in llama.cpp modified weights other or no reference

Rank-1 refusal-direction projection (λ 1.5) on residual writers, then imatrix llama-quantize: routed gate/up IQ4_XS, down IQ4_NL, trunk IQ4_NL, token embd Q8_0; GGUF n-gram IQ4_NL; native .hgn n-gram FP8 E4M3FN from BF16

Reference
BF16 n-gram table (reconstruction error only)
Eval data
seeded sample of 368,640 embedding values across all 128 shards
Tool
not stated
  • No end-to-end quality claims: card states no perplexity, KL-divergence, broad capability benchmark or controlled quality comparison has been run [L151-159]
  • Only figures are n-gram table reconstruction RMSE vs BF16 (weight-level) and pass/fail integration and refusal gates [L100-126, L153-156]
  • Weights modified: abliterated (rank-1 refusal projection, λ = 1.5) [L29, L88-91]
  • Imatrix collected on the base model and applied to edited weights; card calls this an approximation [L82-84]
  • Refusal gate on 12 harmful / 40 harmless prompts with phrase matching; card says it cannot certify a general removal rate [L118-126]

Card snapshot 86d977f7eb14 · updated 2026-10-12

PaoAI/Qwen3.8-Flash-Next-PaoAI-STRIX-BALANCED-2-GGUF

GGUF PaoAI ROCm/HIP engine (pwilkin strix-halo branch + PaoAI fix); previous file on Vulkan fork paoai-strix-engine other or no reference

Organ-by-organ mix: experts gate/up IQ2_XXS (imatrix), experts down IQ4_NL, PLE table IQ4_NL, dense + output head Q8_0 (2.1) / F16 (2); unsloth BF16 source + imatrix

Reference
none (no BF16 run; PPL compares the publisher's two files)
Eval data
wikitext-2 at 8K context (chunks not stated); own coding exam 'chain test' (3 runs, graded by running code); PaoAI Quality Bench v1.3 (55 tasks, one run, temperature 0)
Context
8192
Tool
not stated (perplexity)
  • No BF16 reference: the only PPL figure compares BALANCED-2.1 against BALANCED-2 [L59-60]
  • Card says the −0.16 % PPL change is 'within error' but prints no error bars [L60]
  • Coding exam and Quality Bench are the publisher's own suites with no base-model score for comparison [L64-76, L137-149]
  • Quality Bench is one run per task; 10 of 55 tasks exhausted the token budget and scored 0 [L139, L154-156]
  • Previous-version numbers come from a different engine (Vulkan fork) [L111-123, L292]

Card snapshot 862ba284c8cb · updated 2026-09-27

pentacoxian-dev/Qwen3.8-Flash-Next-IQ3E-Q8D-MTP-GGUF

GGUF patched llama.cpp v0.5.0 only (V100 patch v3) other or no reference

Repack: routed experts + PLE from unsloth UD-IQ3_XXS, dense tensors Q8_0 from unsloth UD-IQ4_XS (rev c8b5954a), plus MTP head converted from Qwen checkpoint at Q8_0 and a Q4_0 draft-only LM head

Reference
none (no BF16 run)
Eval data
not stated
Context
4096
Tool
not stated (perplexity)
  • No reference model: PPL compared only across runtime patch versions/split modes, not against BF16 [L315-317]
  • PPL dataset not stated; only '4K-token chunks' [L315]
  • Not an original quant: tensors are unsloth's UD-IQ3_XXS/UD-IQ4_XS [L31-32, L39]
  • Other quality evidence is a 3-case smoke test [L312]
  • Headline numbers are decode speed (2x V100), not quality [L250-254]

Card snapshot 75e6411f0157 · updated 2026-09-27

perkel/Qwen3.8-Flash-Next-MC

MegaCapybara (repacked EXL3) MegaCapybara (next release; current 1.35 cannot read them) vs BF16

Repack of turboderp's EXL3 quantization (dense/head 4-bit; routed experts 2-bit Tiny, 3-bit Small from turboderp 3.05 bpw); MTP layer re-quantized at 4 bits from BF16

Reference
Qwen's BF16 weights computed with FP32 math
Eval data
held-out set (81,880 tokens: reasoning, math, code, chat, wiki); chat/tool calls written by Flash-Next; real agentic coding sessions
Tool
MegaCapybara engine
  • Not an original quant: routed/dense weights are turboderp's EXL3, repacked [L28-31]
  • Files not yet loadable by the current MegaCapybara release [L34-35]
  • Sizes 65/80 GB include shared trunk, n-gram, MTP and vision files; routed-expert files alone are 30.7/45.8 GB [L41-50]
  • Token counts given only for the held-out set [L66-68]

Card snapshot 19407055c34a · updated 2026-10-12

petergilani/Qwen3.8-Flash-Next-MLX-6bit

MLX mlx-lm 0.31 + bundled ngram_tier.py (n-gram table streamed from paired repo) reference not stated

Routed experts 6-bit g64, dense/attention linears 8-bit; vision and MTP kept; BF16 n-gram table externalized to petergilani/Qwen3.8-Flash-Next-bf16-ngram

Reference
not stated
Eval data
not stated
Tool
not stated
  • Single perplexity figure with no dataset, reference, context or token count stated [L45-46]
  • 106.52 GB excludes the 102.4 GB BF16 n-gram table shipped in a separate repo [L18, L21, L32-33]

Card snapshot ce5f824d4ada · updated 2026-09-08

pfeifferj/Qwen3.8-Flash-Next-GSQ-RCO-GGUF

GGUF llama.cpp upstream f3f1a8f vs BF16

Community reproduction of GSQ + RCO (not IST-DASLab): GSQ on routed experts 8 steps / 64 train seqs; RCO search kept initial allocation; separate embedding shard (BF16 or Q4_0/Q8_0 copied from original quantized release)

Reference
BF16 reference (own run, matching layer placement)
Eval data
MMLU-Pro 2,000 stratified questions (zero-shot, logprob of single-token answers, no chat template); PPL on eight 1,024-token contexts / 4,088 scored tokens of document-disjoint held-out text; IFEval 16 items; GSM8K 8 items
Context
2048
Tool
llama.cpp + own scorer (eval package)
  • Quant scores 2.85 pp above BF16 on MMLU-Pro (McNemar p=0.0048); a gain from 3-bit quantization is unusual and may come from the eval setup rather than the weights [L82-85]
  • Tiny PPL/KL eval: 4,088 scored tokens; IFEval 16 items, GSM8K 8 items [L91, L95-96]
  • KL is approximate, from a uint16 reference cache [L98]
  • All evals use the BF16 embedding shard (151.62 GB combined, derived); the 77.42 GB Q4_0-embedding configuration is unevaluated [L76, L120]
  • RCO search: 'No step improved on step 0, so the initial allocation was kept' — RCO effectively not applied [L146]
  • File named 3.5bit but main file is 3.065 bpw; 3.500 only with Q4_0/Q8_0 embeddings, 6.855 with BF16 [L59, L67]
  • Third-party reproduction, not an IST-DASLab release [L36]

Card snapshot df4f5bbd0a93 · updated 2026-09-23

pipenetwork/Qwen3.8-Flash-Next-MLX-4bit

MLX mlx-lm + bundled qwen4_exp.py (trust_remote_code) vs BF16

MLX affine 4-bit everywhere (experts/dense g64, n-gram tables g32), routers etc. bf16

Reference
bfloat16 (upstream), same runtime and windows
Eval data
wikitext-2 (test), 296,815 tokens in 145 windows of 2048
Context
2048
Tool
bundled mlx qwen4_exp.py runtime; paired per-window ΔNLL, bootstrap 20,000 resamples
  • Uses wikitext-2 PPL; agentionai's AP card reports wikitext is memorised by this architecture's n-gram table
  • MTP head not included in the build [L24]
  • Runs only with the repo's bundled runtime (open, unmerged mlx-lm PR plus three fixes) [L29-30, L42-49]
  • Publisher calls this uniform 4-bit build 'dominated' and published for completeness: +20.6% PPL, worse on 145/145 windows [L83, L89]

Card snapshot aa7c790e804b · updated 2026-08-26

pipenetwork/Qwen3.8-Flash-Next-MLX-6bit

MLX PipeNetwork MLX port (qwen38-flash-next-mlx) vs BF16

Uniform 6-bit affine: experts and non-experts group 64, n-gram tables group 32; routers/gates/vision BF16

Reference
bfloat16 (upstream), identical windows through the same runtime
Eval data
wikitext-2 test, 296,815 tokens in 145 windows of 2048; paired per-window ΔNLL bootstrapped (20,000 resamples)
Context
2048
Tool
own MLX runtime
  • Only wikitext-2 PPL; no held-out or task benchmarks
  • 4-bit ablation builds (not published) are cited in a separate table [L93-101]

Card snapshot b4ccbdba8cd3 · updated 2026-08-26

pipenetwork/Qwen3.8-Flash-Next-MLX-8bit

MLX mlx-lm + bundled qwen4_exp.py (trust_remote_code) vs BF16

MLX affine 8-bit: experts and dense g64, n-gram tables g32, routers etc. bf16

Reference
bfloat16 (upstream), same runtime and windows
Eval data
wikitext-2 (test), 296,815 tokens in 145 windows of 2048
Context
2048
Tool
bundled mlx qwen4_exp.py runtime; paired per-window ΔNLL, bootstrap 20,000 resamples
  • Uses wikitext-2 PPL; agentionai's AP card reports wikitext is memorised by this architecture's n-gram table
  • MTP head not included in the build [L24]
  • Runs only with the repo's bundled runtime (open, unmerged mlx-lm PR plus three fixes) [L29-30, L42-49]
  • 95% CI straddles zero; card calls 8-bit statistically indistinguishable from bf16 on this corpus [L80, L89]

Card snapshot f1c23d4870dd · updated 2026-08-26

pipenetwork/Qwen3.8-Flash-Next-MLX-mixed-4_8bit

MLX mlx-lm + bundled qwen4_exp.py (trust_remote_code) vs BF16

MLX affine: routed experts 4-bit g64, n-gram tables 4-bit g32, attention/DeltaNet/hyper-connections/shared experts/embeddings/lm_head 8-bit g64, routers etc. bf16

Reference
bfloat16 (upstream), same runtime and windows
Eval data
wikitext-2 (test), 296,815 tokens in 145 windows of 2048
Context
2048
Tool
bundled mlx qwen4_exp.py runtime; paired per-window ΔNLL, bootstrap 20,000 resamples
  • Uses wikitext-2 PPL; agentionai's AP card reports wikitext is memorised by this architecture's n-gram table
  • MTP head not included in the build [L24]
  • Runs only with the repo's bundled runtime (open, unmerged mlx-lm PR plus three fixes) [L29-30, L42-49]
  • 95% CI excludes zero: +0.0128 nats/token, worse than bf16 on 128 of 145 windows [L82]

Card snapshot b2c422f3c643 · updated 2026-08-26

primitive-ai/Qwen3.8-Flash-Next-mixed-NVFP4-FP8

NVFP4+FP8 vllm/vllm-openai:qwen38-flash-next image (stock; VLLM_GDN_DECODE_KERNEL=triton required) other or no reference

Routed experts NVFP4 g16; QSA q/k/v/o and GDN in_proj/out_proj FP8 E4M3 per-channel; weights-only round-to-nearest, no calibration; n-gram table BF16; MTP/vision/lm_head BF16

Reference
none (no BF16/FP8 base run on card)
Eval data
1,170-item knowledge suite + 200-item tool-calling suite (same as plain NVFP4 card)
Tool
own harness, auto-scored
  • No base-model (BF16/FP8) reference run [L318-328]
  • Knowledge is a single run (n=1) while the comparison section claims means of at least three runs per build [L48, L363]
  • nvidia cross row uses primitive's INT4 PLE table and no MTP — not nvidia's checkpoint as shipped [L286-287]
  • Cross table and comparison prose are copied verbatim from the plain NVFP4 card (same sitting) — not independent evidence
  • HF listing (190.99 GB) includes the optional mtp_fp8/mtp_nvfp4 overlays; card 183.7 GB excludes them

Card snapshot 07915ee79ec2 · updated 2026-09-06

primitive-ai/Qwen3.8-Flash-Next-NVFP4

NVFP4 vllm/vllm-openai:qwen38-flash-next image (stock) other or no reference

Routed experts NVFP4 g16, weights-only round-to-nearest, no calibration; n-gram table BF16 (FP8 release values); everything else BF16

Reference
none (no BF16/FP8 base run on card)
Eval data
1,170-item knowledge suite (MMLU-Pro, ARC-C, HellaSwag, WinoGrande, CSQA, BoolQ, OBQA, GSM8K, MATH-500) + 200-item tool-calling suite (BFCL v4, xLAM, ToolACE, Glaive v2, When2Call)
Tool
own harness, auto-scored
  • No base-model (BF16/FP8) reference run: only comparisons among quants [L281-291]
  • Section says 'Both halves are means of at least three runs per build' but knowledge here is n=2 (mixed sibling n=1) [L338 vs L291]
  • nvidia cross row is nvidia's experts with primitive's INT4 PLE table and no MTP — not nvidia's checkpoint as shipped [L249-250]
  • Per-benchmark detail table is the mixed sibling's sitting (MMLU-Pro 89.0, GSM8K 97.0), not this repo's (90.0 / 98.0) [L351, L371-373 vs L46]
  • Tool-calling single runs carry ±1.5 spread; most cross gaps are ties per publisher [L219]

Card snapshot e1b0da67618c · updated 2026-09-06

pugant/Qwen3.8-Flash-Next-ROCMFP4_STRIX_LEAN-GGUF

GGUF ROCmFPX llama.cpp fork + pugant/strix-nebulosa lab build (stock llama.cpp cannot load) other or no reference

ROCmFP4 STRIX_LEAN (ftype 106) from unsloth BF16 export with full unsloth imatrix, PLE Q5_1; plus FP2/FP4/Q3 deterministic 2-bit mix (v2) with Q5_1 disk-resident PLE

Reference
none (no BF16 run; FP2 build expressed as ratio to own LEAN build)
Eval data
Italian holdout (Dante, memorized by the base model) and English calibration corpus (agentic / tool-call prose), full corpora
Context
8192
Tool
llama-perplexity (Vulkan build of the fork, -fa on, --no-mmap)
  • English eval corpus is the calibration corpus (evaluation on calibration-domain data) [L126, L229]
  • Italian holdout is memorized by the base model; publisher says it is a pipeline check, not a quality benchmark [L128]
  • No BF16 reference; LEAN labelled 'full quality' with no BF16 comparison [L41, L53]
  • Task-level behaviour of the v2 2-bit build not re-measured [L223, L238]
  • 2-bit build: 75.87 GiB on disk vs 55.94 GiB measured host peak (PLE disk-resident) [L31-32, L257]

Card snapshot a41ebff2581c · updated 2026-09-04

qtum/Qwen3.8-Flash-Next-GGUF

GGUF llama.cpp vs BF16

imatrix quants from BF16; low tiers 'layered' (PLE table Q4_0, saved bits spent on expert gate/up); router/GDN/hc F32, shared expert Q8_0

Reference
BF16 master (own run)
Eval data
wikitext-2 test, 12 chunks
Context
512
Tool
llama.cpp (perplexity)
  • Tiny eval: 12 chunks x 512 tokens (~6K tokens) [L90]
  • Publisher says these PPLs are comparable only within this table [L91-92]
  • '~14% PPL improvement' of layered vs uniform tiers and 'Q8_0 -> Q4_0 PLE costs nothing' are stated without the comparison rows [L50-53]
  • Card's numbers match 6block/Qwen3.8-Flash-Next-GGUF's card (9 of 9 claim lines identical, detected automatically); treated as copied, not as an independent measurement

Card snapshot ee96c021084f · updated 2026-08-31

Quat3rnion/halogen-qwen3.8-flash-next-v2-abliterated

HGN (Halogen) halogen-flash-server only (Strix Halo gfx1151) modified weights other or no reference

Current file: native v2 conversion of orcarouter's abliterated BF16 (RTN NF4 + trellis, bf16-to-hgn); historical r2: peonist's v2 checkpoint with 149 abliterated tensors re-encoded (LDLQ on WikiText-2 train)

Reference
stock peonist v2 quant (not BF16); applies to historical r2 only
Eval data
WikiText-2 raw test, 70,774 tokens, top-128 distribution
Tool
halogen engine ppl mode
  • Weights modified: abliterated (orcarouter refusal-direction edit) [L5-7, L105-108]
  • Card says no perplexity, refusal or speed result is claimed for the current native v2 file; all quality numbers are for the historical r2 build [L82-83, L100-103]
  • HT43 variant has no quality measurements [L246-250]
  • Historical KL/PPL reference is peonist's 4-bit v2, not BF16 [L367]
  • r2 calibrated on WikiText-2 train and evaluated on WikiText-2 test; card says this likely explains r2 landing below stock [L379-381]

Card snapshot 2fd96054e10f · updated 2026-10-13

r0b0tlab/Qwen3.8-Flash-Next-EXL3-2.50bpw

EXL3 ExLlamaV3 v1.5.0 with MoE CPU offload other or no reference

ExLlamaV3 v1.5.0 convert.py -b 2.50 -mb 4 -vb 6 -hq -ngb 3: experts/decoder 2.50 bpw, hq attention/shared, vision 6 bpw, MTP 4 bpw, n-gram table 3 bpw

Reference
none (no BF16 comparison)
Eval data
frozen Q200v2 kit, 180 rows: gsm8k 80, hard_reasoning 20, humaneval 40, ifeval 39(+1); NIAH at 262,080 tokens
Tool
own harness (github.com/r0b0tlab/qwen38-flashnext-exl3)
  • No BF16 or other reference; absolute pass counts on small subsets (20–80 items per family) [L124-132]
  • Benchmark families are subsets (Q200v2 kit), not full benchmarks [L124]
  • Size 61 GiB (64.6 GB) includes 18.5 GiB n-gram table; quantized weights 41.5 GiB [L35-37]

Card snapshot 61a1a139ef24 · updated 2026-09-17

r0b0tlab/Qwen3.8-Flash-Next-NVFP4-W4A16-sm121

NVFP4 W4A16 (ModelOpt) SGLang (source 7d2e8fcb), 2x GB10 TP=2 other or no reference

Self-quantized ModelOpt NVFP4 W4A16, BF16 activations, BF16 mtp.* NEXTN tensors; FP8 E4M3 KV at runtime

Reference
none (absolute pass counts, no BF16 run on the card)
Eval data
Q200-v2: GSM8K 80, HumanEval 40, IFEval 40, hard reasoning 20, BFCL v4 multi_turn_base structural-hard20 subset
Context
262144
Tool
strict full-context AR profile, native thinking low effort; bfcl-eval 2025.12.17 for BFCL
  • Own composite suite (Q200-v2) with no BF16 baseline on the card [L59-71]
  • BFCL figure is a 20-case subset, not a full-category score [L73-76, L138]
  • Quality measured on the strict AR profile; public run uses a different NEXTN profile [L44-47]
  • No checkpoint size stated (131 shards) [L24]

Card snapshot 8b73dc60cfa9 · updated 2026-08-29

RadixArk/Qwen3.8-Flash-Next-NVFP4

NVFP4 SGLang vs BF16

NVFP4 MoE experts; calib 128 cnn_dailymail articles @512

Reference
BF16 runs on an earlier checkpoint revision
Eval data
GSM8K (1319), AIME26 (30x8)
Tool
own
  • BF16 baselines from an earlier checkpoint revision [L123]
  • Card notes long agentic generations run longer than BF16 [L131]

Card snapshot 7b719225242a · updated 2026-08-26

rapid-mlx/Qwen3.8-Flash-Next-4bit

MLX Rapid-MLX 0.13.2 other or no reference

Mixed affine: PLE 4-bit g32, MoE routing gates 8-bit g64, rest 4-bit g64

Reference
none (compared with rapid-mlx/Qwen3.8-27B-4bit-MTP-MLX, a different model; no bf16 baseline)
Eval data
first N examples: MMLU-Redux 2.0 (228), HumanEval instruct (100, chat-safe local variant), GSM8K 0-shot CoT (100), IFEval (100); thinking off, temperature 0, single run
Tool
lm_eval 0.4.12 via Rapid-MLX 0.13.1 endpoint
  • No bf16 baseline; card states quantization fidelity versus bf16 is not measured [L214-216]
  • Comparison model is a different (dense 27B) model, also 4-bit [L166-168]
  • Sampled subsets (N=100; MMLU-Redux 228); card says 1-4 point gaps are not significant [L208-209]
  • GSM8K harness score affected by last-number extraction; re-score reported beside it [L196-200]
  • ~105 GB download; ~103–105 GB MLX active memory while serving [L46]

Card snapshot afeb96d530bf · updated 2026-10-13

RedHatAI/Qwen3.8-Flash-Next-NVFP4

NVFP4 (compressed-tensors) vLLM vs BF16

LLM Compressor NVFP4 W4A4 on MoE experts only, rest BF16; calibration 1024 samples of open-perfectblend

Reference
Qwen/Qwen3.8-Flash-Next (unquantized), same table
Eval data
GPQA Diamond, MMLU-Pro, AIME 2026, LiveCodeBench, Terminal-Bench 2.1, SWE-bench Verified, SWE-bench Pro, DeepSWE 1.1
Tool
Inspect with vLLM server, single seed
  • Single-seed evaluations [L57]
  • Recovery is capped at 100%, so above-baseline scores (e.g. GPQA 92.9 vs 90.4, AIME 100 vs 96.7) count as 100 [L138]
  • Headline 'highest accuracy recovery' 99.1% vs 97.6% / 98.0% [L56]; the gap is largest on DeepSWE 1.1 (94.3 vs 88.0 / 85.9) [L199-201]
  • Card does not state sample counts, context length or harness settings per benchmark
  • No checkpoint size stated

Card snapshot c8f2fb1b9869 · updated 2026-10-05

Robot-Haus/Qwen3.8-Flash-Next-MLX-oQ3-MTP

MLX oMLX (qwen4_exp + native MTP) other or no reference

oQ3: 3-bit affine g32 base with 746 sensitivity-guided 4/5/6/8-bit overrides, from official BF16

Reference
none
Eval data
sample of 30 per benchmark (MMLU, TruthfulQA, HumanEval, MBPP), no-think mode
Tool
oMLX 0.6.4 built-in benchmarks
  • Card header itself labels the figures 'Anecdotal Intelligence Review' [L27]
  • 30 questions per benchmark, no reference run [L140-146]
  • Card names the repository Vontra/Qwen3.8-Flash-Next-MLX-oQ3-MTP and its download command fetches Vontra's repo, not Robot-Haus [L66, L95, L198]
  • Size given as 92.505 GB / 86.152 GiB; peak memory 64.60–69.42 GB in M1 Ultra runs [L76, L126-131]

Card snapshot 9ad88db89cd4 · updated 2026-09-18

Robot-Haus/Qwen3.8-Flash-Next-oQ3.5e-fp16-mtp

MLX oMLX 0.7.0rc1 other or no reference

oMLX oQ 3-bit mixed precision, group size 64, FP16; custom imatrix corpus (3,629 samples) extending oMLX stock mix

Reference
none (no BF16 or base-model comparison)
Eval data
partial samples: MMLU 1000/14042, HellaSwag 300/10042, TruthfulQA 300/817, HumanEval full 164, MBPP 200/500, LiveCodeBench 100/1055, SafetyBench 300/11435; no-thinking
Tool
oMLX 0.7.0rc1 (M1 Ultra 128GB)
  • No reference: absolute scores only, no BF16 or base-model comparison on the same harness [L167-179]
  • Partial samples (e.g. LiveCodeBench 100 of 1055, HellaSwag 300 of 10042), self-reported, no-thinking [L179, L387]
  • Card gives no file size; peak-memory figures (~94–112 GB) are speed-run measurements [L153-158]
  • Name says 3.5e; card's quantization details say 3-bit group size 64 [L436-437]

Card snapshot 612c2b7889c9 · updated 2026-10-03

Robot-Haus/Qwen3.8-Flash-Next-oQ4e-fp16-mtp

MLX oMLX 0.6.4 / 0.7.0.dev2 other or no reference

oQ4e (oMLX 0.6.4) 4-bit affine g64 with 5-bit boosts, PLE 5.00 bits, imatrix 1024x512 oqe_code_multilingual; fp16 non-quant tensors

Reference
none
Eval data
multi-turn agentic battery: 3 scenarios, 9 tool-call turns
Tool
oMLX on M2 Ultra
  • No PPL/KLD or benchmark numbers; only a 3-scenario agentic smoke test and spot checks [L95-97]
  • Card heading names evsinlb/Qwen3.8-Flash-Next-oQ4e-fp16-mtp and credits the quantization to user evsinlb [L14, L18]
  • Front matter lists license apache-2.0 [L2]
  • Card's numbers match evsinlb/Qwen3.8-Flash-Next-oQ4e-fp16-mtp's card (29 of 30 claim lines identical, detected automatically); treated as copied, not as an independent measurement

Card snapshot c220aa3212a3 · updated 2026-09-20

Sawfwair/Qwen3.8-Flash-Next-MLX-Activation-3bit

MLX mere.run / mlx-vlm (Qwen4Exp-aware) other or no reference

Routed-expert banks fresh MLX affine Q3/group-64 from BF16, scales/biases refit on image+text expert-input second moments; other matrices Q4, n-gram Q4/g32

Reference
publisher's pinned published Q4 checkpoint (mere.run), not BF16
Eval data
61 cases scored by exact expected output + 16-case sealed holdout (incl. 8 image/OCR cases)
Tool
mere.run on 128 GiB Apple Silicon Mac, MTP disabled
  • No BF16 comparison; card states the bounded suites 'don't establish general model quality or BF16 parity' [L75-76]
  • Reference is the publisher's own Q4 checkpoint; the three misses matched Q4 output [L68]
  • Memory is holdout peak footprint (19.58% below Q4) [L71-72]

Card snapshot c699bd611366 · updated 2026-08-29

Sawfwair/Qwen3.8-Flash-Next-MLX-Activation-3bit-Native-PLE

MLX mere.run (> v0.45.0) / mlx-vlm other or no reference

Routed experts fresh MLX affine Q3 g64 from BF16 with scales/biases refit against image+text expert-input second moments; other core/MTP/vision Q4; n-gram table Q4 g32

Reference
publisher's pinned published Q4 checkpoint (no-regression comparison), not BF16
Eval data
61 cases by exact expected output + sealed 16-case holdout (8 image/OCR)
Tool
mere.run on 128 GiB Apple Silicon Mac, MTP disabled
  • Reference is the publisher's own Q4 checkpoint, not BF16; card itself says the bounded suites don't establish general quality or BF16 parity [L73-74, L85-86]
  • Small evals: 61 cases + 16-case holdout [L77, L79-80]
  • Memory figures are holdout peak footprint [L81-82]

Card snapshot 1cee9301c745 · updated 2026-08-31

scorpoon/Qwen3.8-Flash-Next-exl3-SC

EXL3 ExLlamaV3 >= v1.5.3 / v1.5.4 / v1.6.0 (per build) vs BF16

ExLlamaV3 convert.py with sc_optimize bit allocation; 8-bit output head; separate 6-bit n-gram table (not uploaded); calibration 217–467 x 2048 rows of agentic / self-sampled traces

Reference
Qwen3.8-Flash-Next BF16
Eval data
held-out evaluation trace (~45k tokens, disjoint from calibration data)
Tool
not stated (ExLlamaV3)
  • Sizes are weight shards only; the 36.4 GiB n-gram table is not uploaded and must come from turboderp's repo [L29-33, L51]
  • Eval is a single ~45k-token trace of unstated content; tool/context not stated [L55]
  • Calibration for 3.15bpw is agentic traces and for 2.05bpw in-domain self-sampled traces; eval is described as disjoint from calibration [L39-43, L55]

Card snapshot 81652966b01e · updated 2026-10-09

ScottzillaSystems/Qwen3.8-Flash-Next-CYBERSECURITY-NVFP4

NVFP4 SGLang (modelopt_fp4), 2x DGX Spark modified weights other or no reference

Refusal removal (weight-level edit, 'CRACK') on Qwen3.8-Flash-Next, NVFP4; quantization recipe not stated

Reference
'base (NVFP4)' — an unnamed NVFP4 build, not BF16
Eval data
MMLU-logit n=40/subject (2280 questions); hard multi-step questions (count not stated)
Tool
not stated
  • Weights modified: abliterated/refusal-removed model [L29, L43-49]
  • Baseline is an unnamed NVFP4 'base', so these numbers measure the abliteration, not quantization loss [L132]
  • No size or quantization recipe stated
  • Per-subject MMLU deltas up to −22.5 pp at n=40 per subject [L140]
  • Card's numbers match dealignai/Qwen3.8-Flash-Next-CYBERSECURITY-NVFP4's card (82 of 82 claim lines identical, detected automatically); treated as copied, not as an independent measurement

Card snapshot b62c93b63411 · updated 2026-09-11

sh0wie/Qwen3.8-Flash-Next-REAP-288-GGUF

GGUF llama.cpp / Ollama / LM Studio modified weights other or no reference

REAP expert pruning 512 -> 288 per layer (calibrated on ~686K tokens of agentic-coding traffic), then single-step GGUF quantization from the pruned bf16

Reference
unpruned full model (precision not stated)
Eval data
HumanEval (164), single run, no CIs
Tool
not stated
  • Card gives HumanEval as 91.5% and as 149 of 164; 149/164 is 90.85% [L25]
  • Pruned: 224 of 512 experts per layer removed — not a faithful quant
  • Only quality number (HumanEval) is measured on the MLX 4-bit lineage, not on any of these GGUF files [L25-28]
  • Single-run, no confidence intervals; calibration is agentic-coding only, general-domain retention not claimed [L73-78]
  • Card 'GB' sizes 78/87/116 match HF listing in GiB (78.05/87.37/115.75); in GB they are 83.81/93.81/124.29 [L43-45]

Card snapshot 2af8d948a587 · updated 2026-08-29

sh0wie/Qwen3.8-Flash-Next-REAP-288-MLX-4bit

MLX mlx-vlm (qwen4_exp MTP support) / omlx >= 0.6.4 modified weights other or no reference

REAP expert pruning 512 -> 288 per layer (top-10 routing kept) of Sawfwair/Qwen3.8-Flash-Next-MLX-4bit (affine 4-bit g64, n-gram g32), calibrated on-device over ~686K tokens of agentic-coding traffic; RMSNorm re-centered, n-gram tensors renamed

Reference
stock 512-expert 4-bit conversion (same harness), not BF16
Eval data
HumanEval 164 problems, unit-test verified, one run per build; rare-name sampling 10 seeded generations
Tool
not stated (same harness, same machine)
  • Weights modified: pruned (224 of 512 routed experts removed per layer) [L23]
  • Baseline is the publisher's 512-expert 4-bit conversion, not BF16 [L20, L112]
  • Single-run HumanEval (164 problems), no confidence intervals; card itself says 1–2 point differences are within noise — 288 (91.5%) scores above 320 (90.9%) [L118-119, L210-211, L114-115]
  • KL divergence against the stock model mentioned for the ladder but no values given [L101-102]
  • Calibration from one team's agentic-coding traffic; card warns non-code domains may degrade more [L207-209]
  • Only text quality evaluated; vision untested after pruning [L37-38, L215]
  • Resident ~68 GB, or ~39 GB with n-gram table streamed [L21, L29-30]

Card snapshot 668f31bcc56b · updated 2026-08-29

sh0wie/Qwen3.8-Flash-Next-REAP-288-Q8E-MLX

MLX mlx-vlm (stock, qwen4_exp MTP support) / omlx v0.6.4+ modified weights other or no reference

REAP expert pruning 512->288 per layer (calibrated on ~686K tokens of agentic-coding traffic, on quantized weights) of Sawfwair MLX 4-bit; surviving expert tensors re-quantized to 8-bit g64; backbone 4-bit; norm/name fixes

Reference
Base Q4 (512 experts, Sawfwair/Qwen3.8-Flash-Next-MLX-4bit) and plain 4-bit prune; no BF16
Eval data
HumanEval pass@1; rare-name sampling reliability (10 trials)
Tool
not stated; single-run, no confidence intervals
  • Pruned: 288 of 512 experts per layer kept (REAP) [L27-28]
  • Card states the 8-bit experts scored at or below the plain 4-bit prune and recommends the smaller build [L44-49]
  • Single-run evaluations, no confidence intervals; card says 1-2 point differences are within noise [L151-152]
  • Calibration on one team's agentic-coding distribution; only text quality evaluated [L39-40, L148-150]
  • Resident ~100 GB with table in memory, 70 GB when n-gram table streamed (omlx) [L21-25, L112-114]

Card snapshot e3e7da4b9556 · updated 2026-08-29

sh0wie/Qwen3.8-Flash-Next-REAP-384-MLX-4bit

MLX mlx-vlm (git main after 2026-08-27); NVMe n-gram mode needs publisher's unreleased patch modified weights other or no reference

REAP expert pruning 512 -> 384 per layer (saliency calibrated on the quantized weights over ~686K tokens of agentic-coding traffic) applied to Sawfwair/Qwen3.8-Flash-Next-MLX-4bit (affine 4-bit g64, n-gram table g32); RMSNorm re-centering and n-gram tensor renames

Reference
stock 512-expert MLX 4-bit conversion (same harness), not BF16
Eval data
HumanEval, 164 problems, unit-test verified, one run per build
Tool
not stated; same harness and machine for all builds
  • Weights modified: REAP expert pruning (384 of 512 experts kept) [L26-27, L117]
  • Single-run HumanEval (164 problems), no confidence intervals; card says 1-2 point differences between neighbours are within noise [L98-99, L134-135]
  • Reference is the stock 4-bit conversion (identified in provenance as Sawfwair/Qwen3.8-Flash-Next-MLX-4bit), not BF16 [L20, L115]
  • Calibration is one team's agentic-coding traffic; card warns non-code domains may degrade more [L118, L131-133]
  • 51 GB resident requires the publisher's NVMe n-gram patch; ~80 GB resident when loaded fully [L23-24, L107-110]
  • Card mentions an eleven-point ladder but shows five rows; 320 scores below 288 on the card's own table [L87, L92-96]

Card snapshot c9bb01bda259 · updated 2026-08-27

showkieo/Qwen3.8-Flash-Next-Sushi-3bpw

EXL3 (sushi pack) sushi v1.0.0+ (Apple Silicon, macOS 26.2+) only vs BF16

Routed experts EXL3 (own converter); dense 8-bit affine; n-gram table 4-bit affine g32 read from SSD

Reference
bf16 checkpoint
Eval data
16 prompts x 512 tokens, scored to the first EOS (7186 positions)
Tool
sushi (same build for every pack, KV kv8)
  • Card's download commands point to beamster/Qwen3.8-Flash-Next-Sushi-3bpw and beamster/...-Sushi-4bpw, not this repo [L44, L144]
  • Card text appears to be beamster's Sushi-3bpw card: the packs labelled as the publisher's own are beamster's repos [L44, L113-120, L144]; treated as a copy of beamster's card, not as an independent measurement
  • Its numbers differ from beamster's current card for the same packs (Sushi-3bpw 0.1036 / 90.31%, Sushi-4bpw 0.0592 / 92.89% there); beamster's card may have been re-measured since
  • Small eval: 16 prompts, 7186 scored positions [L110]
  • Sizes are GPU-resident weights; n-gram table excluded; disk is ~79 GiB [L32, L111]
  • Packs measured with different n-gram tables [L125-126]

Card snapshot 851e86b8d19b · updated 2026-09-27

SiliconSpecies/Qwen3.8-Flash-Next-EXL3-Sushi

EXL3 (Sushi pack) Sushi (Apple Silicon); -hq speeds need a dense hyper-connection read engine patch vs BF16

Routed experts re-encoded in EXL3 trellis format by publisher's own converter at 5 or 6 bits; 8-bit trunk/embeddings/head/MTP from beamster/Qwen3.8-Flash-Next-Sushi-4bpw; -hq keeps hyper-connection mixes bf16; n-gram table is the base model's own (95 GiB) assembled by script

Reference
bf16 teacher (whole model)
Eval data
wikitext 16 x 512 tokens; 16 real reasoning traces; 4 x 32,768-token documents
Context
512 (wikitext) / 32,768 (32K set)
Tool
not stated; 95% bootstrap intervals
  • Small eval sets: wikitext 16 x 512 tokens, 16 reasoning traces, 4 x 32K documents [L38]
  • Sizes (~78 / ~92 GiB) exclude the 95 GiB n-gram table, which is downloaded separately from Qwen's shards; 6bpw-hq ~99 GB resident [L24-26, L54-55, L65]
  • Headline 'several times closer to the bf16 model than 4-bit affine packs at similar memory' rests on the 'port' row, whose source repo is not named [L17-18, L32]
  • Note: 6bpw-hq reasoning KLD (0.0099) is higher than 5bpw-hq (0.0095) on the card's own table [L35-36]

Card snapshot 950fcc19cf08 · updated 2026-10-15

sirfyyn/Qwen3.8-Flash-Next-FLE-GGUF

GGUF upstream llama.cpp loads files; speed numbers and MTP head need FLE fork vs BF16

Layer-wise splices of public quants (Unsloth UD-IQ4_XS/UD-IQ3_XXS/UD-Q2_K_XL, ISTA GSQ-RCO IQ3_S/IQ3_XXS/IQ2_XS) chosen by measured per-layer expert error; no new quantization

Reference
BF16 logits (same build and settings)
Eval data
3 public texts: E Python reference (63 KB), W 4 German Wikipedia articles (81 KB), K German Kubernetes docs (47 KB)
Context
2048
Tool
upstream llama-perplexity, llama.cpp a8c9a4e7c
  • Files are splices of other publishers' quantized layers (Unsloth, ISTA-DASLab); card says so [L50-53]
  • Small eval texts: 63 KB, 81 KB, 47 KB; two of three are German [L206-208]
  • Card treats KLD differences below about 0.01 on a single text as no finding (GPU0 vs GPU1 gave 0.0905 vs 0.0975) [L258-260]
  • Task benchmarks (IFBench, GPQA, LCB) ran on the FLE server with the internal re-trained MTP head c2 (not the shipped head), once each, with no BF16 run; card says they are not a measurement of quantization loss [L315-319]
  • FLE engine's expert-cache router bias (0.001) is not reflected in the KLD numbers [L262-264]

Card snapshot 1b64f8e9ca3e · updated 2026-10-10

sirfyyn/Qwen3.8-Flash-Next-FLE-REAP448-GGUF

GGUF upstream llama.cpp (quality); FLE engine for speed/MTP modified weights vs BF16

REAP expert pruning of FLE-IQ3_M (itself spliced from unsloth UD-IQ4_XS and ISTA-DASLab GSQ-RCO): 448 of 512 routed experts kept per layer, byte-copied, no re-quantization; calibration = FLE-IQ3_M's own answers (256 x 2,048 tokens)

Reference
BF16
Eval data
three public texts: E Python language reference, W four German Wikipedia articles, K seven German Kubernetes docs; MKQA 400 questions in German and English (paired, fixed seed)
Context
2048
Tool
upstream llama-perplexity (llama.cpp a8c9a4e7c) for KLD; FLE server for MKQA
  • Pruned: 64 of 512 routed experts per layer removed — not a faithful quant [L22-24]
  • Card notes KLD understates what pruning removes; MKQA English drop −3.2 points is at the edge of its interval [L93, L114-115]
  • KLD texts are small (Python reference, 4 Wikipedia articles, 7 Kubernetes docs); token counts not given on this card [L77-78]
  • Task benchmarks of the main card were run on FLE-IQ3_M only, not this file [L152-153]
  • KLD measured without the FLE engine's router bias toward cached experts [L156-157]

Card snapshot 6bbaae7c5513 · updated 2026-10-10

sleepyeldrazi/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-HCQ5_0

GGUF stock llama.cpp (trunk); sleepyeldrazi/llama.cpp-strix feat/strix-prod for MTP reference not stated

ISTA-DASLab GSQ-RCO IQ3_XXS pack with 194 hyper-connection tensors requantized BF16 -> Q5_0; all other tensors byte-identical; MTP draft sidecars (full and 65,280-token head)

No evaluation method stated.

  • Card states perplexity was not measured for this build; quality gates are a 128-token greedy comparison (identical for first 47 tokens), a 33K needle and a 1K coherence check [L56-62]
  • Base-pack wikitext PPL 3.9044 is quoted from GSQHalo.cpp, not measured by this publisher [L61-62]
  • Trunk shard 46.2 GiB plus 54.4 GiB PLE shard (Q8_0, byte-identical to base pack) [L68-69]
  • Not an original full quant: derived from ISTA-DASLab's IQ3_XXS pack [L27-29]

Card snapshot cb186e74baa7 · updated 2026-10-08

Slomin/Potato-CODER-48GB

GGUF llama.cpp potato branch (stock llama.cpp loads without MTP) modified weights other or no reference

Routed experts pruned 512 -> 256 per layer by routing statistics on a 16-language coding/agentic corpus; IQ4_XS experts with imatrix from the unpruned model; 610 non-expert tensors Q6_K/Q8_0

Reference
full 512-expert model in the publisher's own IQ4_XS (not BF16)
Eval data
20 sealed + 40 development tasks x 4 samples per language family (16 families); pre-registered gate: one-sided 95% upper bound of (full - v1.0) < 15 points
Tool
llama-server (potato branch)
  • Model files withdrawn 11 Oct 2026 after the publisher found degradation incl. repetitive reasoning loops [L36-40]
  • Pruned (half the routed experts removed): not a faithful quant [L70-71]
  • Reference is the publisher's own unpruned IQ4_XS, not BF16 [L84]
  • NVIDIA comparison uses different engines, quantizations, machines, thinking budgets and parallelism, one run per benchmark; card labels it descriptive [L129-130]
  • Prose says 'on competitive coding the pruned model solved more than the full one' while the comparison is against NVIDIA's NVFP4 under a different setup [L125, L129-130]
  • Size is the main file only (60.99 GiB); MTP head 1.3 GiB and projector 0.6 GiB separate [L78]

Card snapshot a96007bb288e · updated 2026-10-11

sorenmat/Qwen3.8-Flash-Next-NVFP4

NVFP4 vLLM (Blackwell B200/B300) other or no reference

Card identical to nvidia/Qwen3.8-Flash-Next-NVFP4 (ModelOpt v0.46.0: routed experts W4A4 NVFP4, attention/shared BF16, MTP experts FP8 128x128, PLE per-tensor FP8)

Reference
Qwen/Qwen3.8-Flash-Next-FP8 (not BF16)
Eval data
GPQA Diamond, HLE, τ²-Bench Telecom, MMMU Pro, SciCode, AA-LCR, IFBench, Omniscience, Terminal-Bench 2.1
Tool
not stated
  • Card is byte-identical to nvidia/Qwen3.8-Flash-Next-NVFP4's card; the evaluation numbers are NVIDIA's, not measured on this repo [L39]
  • Baseline is the official FP8 release, not BF16 [L113]
  • Sampled at temperature 1.0 with no run count or error bars [L113]
  • No absolute size, only '~2.7x smaller than BF16 (about 63% reduction)' [L186]
  • Card's numbers match nvidia/Qwen3.8-Flash-Next-NVFP4's card (6 of 6 claim lines identical, detected automatically); treated as copied, not as an independent measurement

Card snapshot 70232878464c · updated 2026-10-14

spectator2026/Qwen3.8-Flash-Next-W6-HQQ-g32

compressed-tensors int6 (W6A16) patched vLLM (pinned nightly + 7 patches + custom 6-bit Marlin kernel), A100 vs BF16

Routed experts int6 symmetric g32: SmoothQuant fold (α=0.65), MSE-clip init, HQQ proximal scale refit; everything else BF16

Reference
BF16 model
Eval data
wikitext; agentic multi-turn conversations (token counts not stated)
Tool
vLLM (earlier preview build; wikitext re-measured on pinned nightly)
  • Eval set sizes, context length and chunking not stated [L106-113]
  • Main figures measured on an earlier vLLM preview build; only wikitext re-measured on the pinned image (0.0223) [L113]
  • Uses wikitext, which agentionai's AP card reports is memorised by this architecture's n-gram table
  • Size given as ≈224 GB (209 GiB) [L25]

Card snapshot 05848f88f4a5 · updated 2026-10-05

SpeederX/Qwen3.8-Flash-Next-UD-Q2_K_XL-EM-GGUF

GGUF Siliang Engine v0.1.8+ only other or no reference

Expert-major repack of unsloth UD-Q2_K_XL; layer-2 gate/up requantized from BF16 IQ3_XXS -> IQ2_XS with unsloth imatrix

Reference
original unsloth UD-Q2_K_XL layout (not BF16)
Eval data
one 64-token continuation of a fixed prompt (65 positions)
Tool
Siliang Engine, teacher-forced, all experts on GPU
  • Single prompt, 65 positions; publisher says this is not a perplexity or benchmark campaign [L66-67]
  • Reference is the quantized unsloth layout; KL is top-10 only [L64-66]
  • Repack of unsloth's UD-Q2_K_XL; only layer 2 gate/up changed [L42-51]

Card snapshot 2a6e332af808 · updated 2026-09-28

spiritfather/cyberneurova-Qwen3.8-Flash-Next-i1-GGUF

GGUF llama.cpp (master) modified weights other or no reference

imatrix (bartowski calibration_datav3 on the model's own bf16); PLE mixed rungs: trunk Q8_0, routed up/gate at named type, down IQ4_NL; plus flat imatrix quants. Source is cyberneurova's uncensored fine-tune

Reference
Q8_0 (own build), not BF16
Eval data
wikitext-2
Context
512
Tool
not stated
  • Source model is an uncensored/unrestricted fine-tune (cyberneurova), not stock Qwen weights [L18, L38]
  • Reference is the publisher's Q8_0, not BF16; 'near-lossless (+0.28% PPL vs Q8_0)' is relative to Q8_0 [L36, L46, L53]
  • Wikitext-2 at ctx 512; chunk count and tool not stated [L46]
  • Card says IQ1/IQ2 quants are not available for this architecture, yet lists an IQ2_M rung [L38, L48]
  • IQ3_M (109.1 GiB) has higher PPL delta than the smaller IQ3_XXS (+3.11% vs +3.05%) while KLD is lower (0.0729 vs 0.0937) [L49-50]
  • Flat imatrix quants (L59-71) have no quality numbers; sizes in GB vs PLE rungs in GiB

Card snapshot d2d2635dd5ec · updated 2026-09-23

steadfastgaze/Qwen3.8-Flash-Next-MoEspressoV3

MoEspresso package (IQ_K/K-quant MLX) MoEspresso 3.0.0 (Apple Silicon) other or no reference

Routed experts IQ3_K (layers 0-1) / IQ2_K (rest); dense/non-routed Q6_K/Q8_0/BF16; original BF16 PLE tables on SSD; Cache-Prior routing bias at serve time; text-only

Reference
hosted models via OpenRouter (Qwen3.8 Flash xhigh and other vendors' models), not BF16 of this checkpoint
Eval data
frozen 48 LiveBench questions (8 each from 6 categories), salted SHA-256 selection, official scorers
Tool
MoEspresso bounded path: medium thinking, greedy, 223 resident expert slots/layer, Cache-Prior factor 2 / 2 protected routes
  • Quality measured with Cache-Prior routing enabled, which changes selected experts and output; result mixes quantization and routing policy [L118-122]
  • Small custom set (48 questions); card notes it differs from the official LiveBench protocol [L100-102]
  • Reference row is the hosted Qwen3.8 Flash API, not this checkpoint's BF16 source; other rows are different vendors' models with provider defaults [L86-98]
  • Size 145.75 GB includes 102.40 GB original BF16 PLE tables; model safetensors 43.35 GB [L143-145]

Card snapshot 7c830239d6ed · updated 2026-09-24

SuperPUPS/Qwen3.8-Flash-Next-INT8x17-AttnINT4

compressed-tensors (INT8/INT4 pack-quantized) vllm-qfn patched vLLM snapshot (feature/qsa-int8-kv, commit e12683d245) other or no reference

Community W4A16 checkpoint (INT4 g128 experts) with 17 of 48 expert layers re-quantized to INT8 g128 from the official FP8 checkpoint (msegrid scale refinement); attention INT4 g128

Reference
official FP8 checkpoint (weight-level rel-Frobenius / cosine); live gates on serving stack with no reference model
Eval data
1536 expert-matrix pairs per layer (weight-level); 3 golden probes, 10-task battery, 8x determinism, NIAH @128K, 30-min soak (live gates)
Tool
quantization/ scripts (rank_sensitive_layers.py, verify_checkpoint.py); vllm-qfn serving
  • Quality evidence is weight-level (INT8 roundtrip cosine 0.999962 vs FP8 reference; parent INT4 0.9925) plus pass/fail live gates; no perplexity/KLD or benchmark scores vs BF16 [L81, L85-99]
  • Per-layer INT4-error ranking table is for the parent W4A16 experts, used to choose INT8 layers [L62-72]
  • Base W4A16 checkpoint is a community build, not named on the card [L52, L226]
  • Sibling INT8x25 is listed as 'all passed' quality gates with no numbers [L38]

Card snapshot 8d7e6f73b95d · updated 2026-10-15

SuperPUPS/Qwen3.8-Flash-Next-INT8x25-AttnBF16

compressed-tensors (INT8/INT4 experts, BF16 attention) patched vLLM (vllm-qfn snapshot) other or no reference

Community W4A16 base; 25 of 48 expert layers re-quantized to INT8 g128 from official FP8 checkpoint (msegrid scales); attention byte-copied BF16 from FP8 checkpoint

Reference
official FP8 checkpoint (proxy) for per-layer INT4 error and INT8 roundtrip cosine
Eval data
1536 expert-matrix pairs per layer (weight-space); live gates: 3 golden probes, 10-task battery (temp=0), NIAH @128K (112,917 prompt tokens)
Tool
vllm-qfn on 2x CMP 170HX, TP2+EP, MTP k=4
  • Per-layer rel-F/cos ranking table measures INT4 error of the W4A16 base layers used for selection, not this file's INT8 layers [L72-90]
  • No output-level comparison vs BF16/FP8 (no KLD, PPL or benchmark scores); quality evidence is weight-roundtrip cosine and pass/fail gates [L96, L106-114]
  • 'the highest-quality W4-class release we could measure' is stated without a comparison table [L25-26]
  • Wave-1 per-layer metrics were not retained [L69-70]

Card snapshot 6c077c300961 · updated 2026-10-15

tcclaviger/Qwen3.8-Flash-Next-MXFP4-FP8-GPTQ

MXFP4+FP8 tcclaviger/vllm:latest only; AMD Radeon AI PRO R9700 only vs BF16

AA2: activation-aware MXFP4 experts + shared experts (GPTQ full-row, damping 0.01, layer-serial, 2M-token 27-domain calib @8192); FP8 attention and MTP; n-gram table int6 g32 fp16 scales

Reference
BF16 base
Eval data
frozen 994-prompt 27-domain set, disjoint from calibration, measured at the end-of-prefill position (KLD/top-1); PPL dataset not stated
Tool
tcclaviger/vllm image (IronLLM suite for PPL)
  • KLD/top-1 measured at one position (end of prefill) per prompt over 994 prompts — ~994 token positions, and not comparable with per-token llama.cpp KLD tables [L134]
  • PPL dataset, context and baseline PPL not stated [L24]
  • Size: banner says 116 GB total but listed components sum to 111.2 GB (2.7+64.2+41.6+2.7); HF listing is 120.33 GB / 112.06 GiB [L18-23]
  • Runs only on the publisher's vLLM image on AMD R9700 [L17, L119]
  • Numbers at L134 (0.055 / 95.2% vs 0.064 / 93.3%) are roundings of L25-26

Card snapshot 6fe3e7464c18 · updated 2026-10-01

TheDrainFlorist/Qwen3.8-Flash-Next-VQ-2.1bpw

MLX Knurlogic / mlx-lm (custom VQ) vs BF16

Data-free vector quantization (k-means codebooks)

Reference
bf16 teacher, cached top-64 logits
Eval data
3 house corpora (prose, code, literary), 12,288 tokens each
Tool
Knurlogic 0.1.3
Unit
KL in millinats/token (divide by 1000 for nats)
  • KL unit is millinats — not comparable numerically to nat-scale KLD tables without conversion

Card snapshot b5591ded8934 · updated 2026-10-11

TheDrainFlorist/Qwen3.8-Flash-Next-VQ-3.2bpw

MLX Knurlogic >=0.1.3 / mlx-lm (custom VQ model.py) vs BF16

Data-free vector quantization (k-means over weight subvectors, seed 1234); experts and PLE d=4/K=2048, six layers d=2/K=256

Reference
bf16 teacher, cached top-64 logits
Eval data
3 house corpora (prose, code, literary), 12288 tokens per corpus
Tool
Knurlogic 0.1.3 (mlx 0.32.3)
Unit
KL in millinats/token
  • KL unit is millinats — not comparable numerically to nat-scale KLD tables without conversion [L159-160]
  • Sizes are text weights only; VQ builds also carry a 0.84 GiB vision tower and every build ships a 2.1 GiB MTP head [L162-163]
  • Affine comparators are the publisher's own conversions, not community builds [L31-32, L311]
  • Headline says 68.7 GiB while the 2026-09-17 changelog gives 69.5 GiB for the same v2 weights [L21, L87]
  • Prose PPL of this build (2.300) is below the bf16 teacher (2.31); card says to rank by KL, not PPL [L173, L180, L188-192]

Card snapshot 33e10e66b427 · updated 2026-10-11

TheDrainFlorist/Qwen3.8-Flash-Next-VQ-4.4bpw

MLX Knurlogic / mlx-lm (custom VQ) vs BF16

Data-free vector quantization (k-means codebooks)

Reference
bf16 teacher, cached top-64 logits
Eval data
3 house corpora (prose, code, literary), 12,288 tokens each
Tool
Knurlogic 0.1.3
Unit
KL in millinats/token
  • KL unit is millinats
  • Top-1 and perplexity are prose-corpus only; mean KL averages three corpora [L170, L178]
  • Table size is the text weights; the repo also ships a 0.84 GiB vision tower and a 2.1 GiB MTP head [L171-172], which is why the card gives about 104 GB on disk [L235]

Card snapshot c13ef42d2362 · updated 2026-10-11

TheDrainFlorist/Qwen3.8-Flash-Next-VQ-5.5bpw

MLX Knurlogic >=0.1.3 / mlx-lm PR #1788 (custom VQ model.py) vs BF16

Data-free vector quantization (k-means/Lloyd, seed 1234): MoE experts d=2/K=1024, PLE d=8/K=4096, flat allocation

Reference
bf16 teacher, cached top-64 logits
Eval data
3 corpora (prose, code, literary), 12288 tokens each, same positions for every row
Tool
Knurlogic 0.1.3 (mlx 0.32.3)
Unit
KL in millinats/token
  • KL unit is millinats — not numerically comparable to nat-scale KLD tables without conversion [L145-146]
  • Affine comparators are the publisher's own conversions, not community builds [L31-32, L302]
  • Sizes are text weights; card also says 123 GB on disk (plus 0.84 GiB vision tower and 2.1 GiB MTP head) [L148-149, L213]
  • Prose PPL of this build (2.244) is below the bf16 teacher's 2.31; publisher says rank by KL, not perplexity [L163, L166, L175-179]
  • Numbers published before 2026-10-04 were on a runtime with wrong PLE hashing and are not comparable [L149-151]

Card snapshot 7740086b1732 · updated 2026-10-11

tiyuvta/Qwen3.8-Flash-Next-NVFP4

NVFP4 (ModelOpt) tiyuvta serving engine (bring-up, single request; not wired into server) vs BF16

ModelOpt 0.46.0 NVFP4 (W4A4, FP8-e4m3 scales, group 16) on routed experts; calibration 256 x 512 tokens abisee/cnn_dailymail 3.0.0 seed 17; MTP block BF16 grafted byte-exact; everything else BF16

Reference
BF16 parent, same box, same prompts
Eval data
4 fixed probe prompts, last-token logits
Tool
not stated
  • Quality evidence is 4 probe prompts, last-token logits only; argmax differs from BF16 on 1 of 4 ('capital of Australia') [L42-49]
  • Card states the artifact is a single-request bring-up and not ready for multi-session serving [L68-76]
  • No size stated on card

Card snapshot 715444446b0d · updated 2026-09-12

tls20/Qwen3.8-Flash-Next-oQ5e-mtp

MLX (oMLX oQ5e) oMLX other or no reference

oQ (oMLX v0.6.4) 5-bit, group 64, oQe enabled, MTP preserved

Reference
none stated
Eval data
MBPP, HumanEval, MMLU, JMMLU, TruthfulQA, LiveCodeBench (subsets, versions, settings not stated)
Tool
not stated
  • No methodology, sample sizes, settings or baseline for any score [L79-88]
  • LiveCodeBench 59% with version unstated; Qwen's card reports 91.9 on LiveCodeBench v6, so the two are not comparable as given [L88]
  • No disk size stated; only peak memory during throughput tests (84.7–99.6 GB) [L42-62]

Card snapshot 6bffcfdd9b9b · updated 2026-10-01

TracNetwork/mayhem-catalog-Qwen-Qwen3-8-Flash-Next-NVFP4

NVFP4 SGLang (qwen4_exp) vs BF16

RadixArk NVFP4 card text (ModelOpt v0.46.0, routed experts W4A4, calib 128 cnn_dailymail @512)

Reference
BF16 runs on an earlier checkpoint revision
Eval data
GSM8K full 1319 (t0.6, top-p 0.95, max 8192); AIME26 30 problems x 8 (t1.0, max 130k)
Tool
not stated
  • Card is RadixArk's text and calls itself 'a private candidate release' [L19-26]
  • AIME26 measured on a previous revision differing in PLE tables, despite 'this exact checkpoint' heading [L116, L127-128]
  • BF16 baselines from an earlier checkpoint revision [L123-125]
  • Card's numbers match RadixArk/Qwen3.8-Flash-Next-NVFP4's card (11 of 11 claim lines identical, detected automatically); treated as copied, not as an independent measurement

Card snapshot ae294c18fb23 · updated 2026-09-30

turboderp/Qwen3.8-Flash-Next-exl3

EXL3 ExLlamaV3 v1.4.5+ vs BF16

EXL3 plain quants (not self-calibrated); five branches 2.05–6.05 bpw

Reference
full-precision model (D_KL(p_FP ‖ p_quant))
Eval data
self-generated in-domain trace, 14,496 input + 45,588 output tokens
Tool
not stated (repo ships qbench_prompts files)
  • Results exist only as chart images; values here are transcribed from the printed labels and checked against marker positions [L27-28]
  • Eval text is a self-generated in-domain trace, so absolute KLD is much lower than on held-out corpora; e.g. unsloth UD-Q4_K_XL is 0.0084 here vs 0.046893 on unsloth's own table
  • The NVFP4 checkpoint measured is not named; W4A16 and W4A4 are two execution modes of the same weights
  • Chart prints a measured noise floor (mean KLD 0.00249): differences near it are not meaningful
  • The GGUF series are unsloth's UD quants (names match unsloth's files); file sizes are not given, only bpw excluding embeddings and output head

Card snapshot c1def4275dc6 · updated 2026-09-01

txgsync/Qwen3.8-Flash-Next-Dynamic-oQ5e-BF16-PLE-mtp

MLX oMLX (commit 68eb444), MLX 0.32.2 vs BF16

Dynamic oQ4e-based mix (experts Q4/Q5 g64-128, attention/GDN Q8, embeddings Q8) with BF16 PLE, MTP and vision; calibration from converted unsloth imatrix; built on A100

Reference
BF16 source
Eval data
four fixed public document windows, 8,192 scored next-token positions each
Tool
not stated
  • Small eval: 4 windows x 8,192 positions [L87-88]
  • llm-bench results are single runs per condition with large spread (coding 22.20 vs 79.00 between seeds); publisher claims no significance [L128-129, L144-149]
  • Card states this build benchmarked better than an earlier Metal build that had lower PPL/KL; that Metal build's numbers are not tabulated [L103-110]
  • Size includes 102.40 GB BF16 PLE and 5.21 GB BF16 MTP [L76-77]
  • Card reproduces Qwen's base-model tables (not claims about the quant) [L177-]

Card snapshot b62713ae1d38 · updated 2026-09-14

Tylogi/Qwen3.8-Flash-Next-EW-MFQ

MFQ (.mfq) MFQ only (CUDA / Metal) reference not stated

Expert-wise mixed precision: per-expert Gate/Up/Down format and bit-width (NINT scalar or NVQ vector) allocated under a global budget with NatScale contrast-calibrated sensitivity

Reference
not stated ('the reference model')
Eval data
WikiText-2
Tool
not stated
  • Reference model, context, chunk count and tool not stated; only 'WikiText-2' [L74]
  • Unsloth rows equal Unsloth's published docs values (rounded) and link to Unsloth's analysis; the card does not say they were re-measured under the same setup as the MFQ and GSQ-RCO rows [L86-91]
  • Uses wikitext, which agentionai's AP card reports is memorised by this architecture's n-gram table
  • Tiers S4-L and S5-L have no quality numbers [L67-68, L78-85]
  • Weight sizes exclude vision encoder and MTP head; V1-M download is 72.44 GiB with 43.36 GiB resident [L97, L153]

Card snapshot 3cdfba425c44 · updated 2026-10-16

umixer/Qwen3.8-Flash-Next-oQ4e-mix6-mtp

MLX oMLX 0.7.0+ (or qwen4_exp-capable mlx_vlm) other or no reference

Component-wise byte-copy merge of Jundot oQ6e (edge-layer experts L0-3/L44-47, attention/GDN, 6-bit g32 n-gram, MTP, shared, embed/head) and Jundot oQ4e (middle 40 layers' experts); no re-quantization

Reference
none (absolute accuracy; no BF16 or oQ5e figures on card)
Eval data
sampled-300 protocol: MMLU-Pro, MATHQA, MBPP (300 each)
Tool
not stated; thinking on, reasoning effort auto, truncation accounted
  • Card claims accuracy 'equal to or above uniform oQ5e' but publishes no oQ5e (or BF16) accuracy figures for comparison [L22-25, L101]
  • Headline 'accuracy on finished' excludes truncated answers (n=279/282/267 of 300) [L139-146]
  • Not an original quant: byte-identical tensors from Jundot's oQ4e/oQ6e packs [L27-30, L98]
  • v1 figures listed inline are for a previous build with different donors [L32-38, L144-146]
  • Package 116.8 GiB (text elsewhere estimates ≈115 GiB); resident 75.49 GB vs uniform oQ4e 69.53 GB measured on same host [L62, L88-91, L98]

Card snapshot 73d2cd0bc0ac · updated 2026-10-15

unsloth/Qwen3.8-Flash-Next-GGUF

GGUF llama.cpp / Unsloth Desktop reference not stated

Unsloth Dynamic (UD) imatrix quants; PLE/n-gram table kept at 4-bit minimum

Reference
not stated
Eval data
not stated
Tool
not stated
  • Headline says 1-bit is 75GB with top-1 of 80%; the table's 1-bit rows are 72.5 GB / 77.325 and 74.5 GB / 79.691 [docs L11, L96-97]
  • KLD methodology entirely unstated on docs page: no reference model, dataset, context or chunk count [docs L90]
  • HF README carries no quant measurements, only 'Unsloth Dynamic 3.0 achieves superior accuracy & outperforms other leading quants' with no evidence on card [README L14]
  • Headline '80% top-1 accuracy recovery ... 79% less disk' [docs L90] / '1-bit is 75GB ... top-1 accuracy of 80%' [docs L11] is top-1 agreement (79.691 at 74.5 GB, UD-IQ1_M), not recovery; rounded up [docs L97]
  • Docs size table '4-bit 96-114 GB' / '8-bit 200 GB' are memory requirements, not file sizes [docs L37]
  • Q8_0 shows mean KLD 0.026574 and top-1 94.122 — unusually high loss for Q8_0, consistent with an undisclosed reference/eval setup [docs L105]

Card snapshot 766911a6b736 · updated 2026-10-06

Vemloir/Qwen3.8-Flash-Next-sc-exl3

EXL3 exllamav3 >= 1.6.1 (tested TabbyAPI on 1.6.2) vs BF16

EXL3 4.05 bpw from BF16 with self-generated calibration (~1M tokens: agent sessions, RU/EN chats, images), measured expert bit split (layers 0-9 down proj more bits), affinity-weighted per-expert Hessians; dense 6-8 bpw, head and vision 8 bpw, n-gram table 8 bpw

Reference
BF16
Eval data
two traces not used in calibration: NVIDIA Nemotron 3.5 Super post-training conversations (48 rows, 23k scored tokens); multilingual Russian + 9 languages (52 rows, 26k scored tokens)
Tool
exllamav3 qbench, teacher-forced on assistant turns; paired bootstrap over rows
  • Small evals: 23k and 26k scored tokens (48 and 52 rows); only 4 rows per language outside Russian, card calls per-language numbers indicative [L40-41, L86-87]
  • Head is 8 bpw vs 6 bpw in turboderp's quant (+~150 MB outside the bpw budget); card says part of the gain comes from it [L84-85]
  • Card says multiple factors differ at once, so the gain is not attributed to any one of them [L85-86]
  • Size 69 GB is model files excluding the 52 GB n-gram table [L24]

Card snapshot c9e11971e843 · updated 2026-10-16

VnimanieAI/Qwen3.8-Flash-Next-W4A16

compressed-tensors W4A16 vLLM dev build (qwen4_exp), Ampere, --enable-expert-parallel other or no reference

RTN INT4 g128 symmetric (no calibration) on routed experts and attention q/k/v/o; shared expert, MTP, PLE, router, indexer, GDN, embeddings BF16

Reference
no BF16/FP8 run of this model: Qwen3.8-27B INT8 W8A16 (a different model) for the greedy suite; hosted qwen3.8-flash endpoint and DeepSeek V4 Flash 0731 for capability benchmarks
Eval data
50-task deterministic greedy suite (arith, sorting, positional retrieval, JSON); τ²-Bench Airline 50 tasks 1 trial; MMLU_RU 1000-q paired subsample seed 1234; GPQA Diamond 198 q zero-shot CoT; 40-q GPQA paired probe vs cyankiwi AWQ
Tool
vLLM + own harness
  • 'Quality matches the base-family reference' is measured against Qwen3.8-27B INT8 — a different model — on a 50-task smoke suite [L28, L96-104, L212]
  • Publisher's own paired probe has cyankiwi AWQ ahead (26/40 vs 19/40) and the card recommends AWQ for 8x24 GB [L32-37, L181-193]
  • τ²-Bench is a single trial; publisher says 4–6 pp differences are within noise [L135-136]
  • GPQA strict 0.601 reflects 37.9% of answers hitting the 32k budget; card attributes much of the verbosity to the RTN recipe [L166-168, L189-192]
  • Size: 168 GB on disk, ~66 GB GPU-resident; 102 GB PLE tables in host RAM [L27]

Card snapshot 142456d8a06f · updated 2026-09-02

WarlaxZ/Qwen3.8-Flash-Next-NInfer-3090

NInfer container (from GGUF) ashalliants/ninfer-3090 fork, Windows other or no reference

Lossless repack of ISTA-DASLab GSQ-RCO IQ2_XS GGUF into .ninfer + Strata's Q2_0 MTP drafter + BF16 vision tower; nothing re-quantized

Reference
Strata 0.1.38 running the same IQ2_XS GGUF (runtime vs runtime, not BF16)
Eval data
Strata's prompts and Infernix's prompts, 2,557 positions each, teacher-forced
Tool
tools/flash_next/strata_compare.py
  • Compares two runtimes on the same 2-bit weights, not quantization loss; card points to DASLab's card for that [L135-137]
  • First set ΔNLL +0.0064 is 2.4 SE above zero (chat text +0.0158, 3.2 SE) [L133-135]
  • 41.0 GB is the model files only; the 29.1 GB n-gram volume is separate (about 70 GB disk) [L97]

Card snapshot c7dfb9694fc3 · updated 2026-10-12

wemoh/Qwen3.8-Flash-Next-strixw

strixw strixite only (AMD Strix Halo gfx1151) vs BF16

Asymmetric integer group quant: routed experts and GDN input 4-bit g128, other dense projections 8-bit g64, MTP input proj 4-bit g64; n-gram table int8 per-row

Reference
FP32 outputs of the unquantized model
Eval data
242 positions of three reference prompts; terminal-bench-mini core19
Tool
strixite
  • Tiny eval: 242 token positions from three prompts [L93-94]
  • ~116 GiB download includes the 48.3 GiB int8 n-gram table read from SSD; 65.9 GiB is main weights [L48, L52-54]
  • 4.25 bits refers to routed weights only [L84]

Card snapshot b64f3a69070c · updated 2026-10-14

whm0627/Qwen3.8-Flash-Next-177B-A3B-fits64GB-PLElast-GGUF

GGUF llama.cpp with qwen4exp support (Metal) other or no reference

Repack of unsloth UD quants with the PLE/n-gram table moved to the end of the file; nothing requantized

Reference
none (no BF16 or unsloth original compared)
Eval data
GSM8K 40 items, HumanEval 20 items; perplexity dataset not stated
Tool
not stated
  • Not an original quant: same weights as unsloth's UD quants, only tensor order changed [L23]
  • Tiny evals: GSM8K 40 items, HumanEval 20 items; publisher says not tested more deeply [L33-34, L47-48]
  • Perplexity dataset, context and chunk count not stated; no reference [L49]
  • Card calls the model 177B ~A3B, whereas other cards give 6B active [L18]
  • VRAM/wired use ~44–48 GB vs 68–79 GB file (n-gram table paged from disk) [L59-60]

Card snapshot 8945f87bd117 · updated 2026-09-07

wrldsuksgo2mars/Qwen3.8-Flash-Next-EXL3-K4.25-v1

EXL3 (GPTQModel fork) QFlashRT (in development; RTX PRO 6000 Blackwell / DGX Spark); GPTQModel EXL3 backend for eval vs BF16

Calibrated mixed K4/K5 EXL3 routed experts averaging 4.25 bits per expert weight (incl. MTP layer); non-expert tensors original incl. BF16 PLE tables

Reference
original BF16 (same harness, per device)
Eval data
exploratory held-out set: 89 records, 67,477 next-token predictions
Tool
original framework block math with GPTQModel EXL3 projection backend; 48 target layers, MTP disabled
  • Card calls the evaluation exploratory; results include backend numerical effects and are hardware-specific [L48, L59-60]
  • On DGX Spark the EXL3 PPL (3.85601) is lower than the BF16 reference (3.89576) on that device [L57]
  • 4.25 bpw describes routed expert storage only, not the whole checkpoint [L16-20]
  • Dataset of the held-out set not named on card [L48-49]
  • Size: 179,366,664,184 bytes tensor payload; 179,406,710,432 incl. headers (BF16 PLE retained) [L22-24]

Card snapshot 73a050c27b8c · updated 2026-09-07

xrrand/Qwen3.8-Flash-Next-MLX-Serve-4bit-mtp

MLX (mlx-serve layout) mlx-serve 26.9.4+ other or no reference

Hybrid: ddalcu MLX-Serve mixed 4/8-bit base layout (attention/GDN/shared/MTP 8-bit, external ngram_table.bin) with routed-expert bytes from Jundot oQ4e swapped in verbatim; embed_tokens re-quantized 8-bit g64 from BF16

Reference
none (absolute tool-use scores; compared with Jundot oQ4e on oMLX and ddalcu base on mlx-serve)
Eval data
Tool-Eval-Bench v2.7.1, 92 scenarios
Tool
Tool-Eval-Bench, temp 0.2, parallel 2, thinking on, effort medium
  • Single agentic tool-use benchmark; scores given as ranges, number of runs not stated [L42-48]
  • Comparison rows use different engines (oMLX vs mlx-serve) [L46-48]
  • Card does not state whether the Jundot and ddalcu scores were run by this publisher [L44-48]
  • Benchmark box names the model 'xrrand/Qwen-3.6-Flash-Next-mlx-serve-4bit-mtp' [L56]
  • Not an original quant except embed_tokens: tensors from ddalcu and Jundot [L29-40]
  • No total size stated; routed experts ~60 GB, ngram_table.bin 30 GB [L31, L109]

Card snapshot afd39de786fa · updated 2026-10-01

xv0y5ncu/Qwen3.8-Flash-Next-GLQ-trellis-3inst-3bpw-ple4

GLQ vLLM + glq>=0.8.20 plugin other or no reference

GLQ: QTIP trellis-coded quantization (3inst) + RHT + LDLQ, 3 bpw decoder, 4 bpw n-gram table; calibration 128 x 2048 WikiText-2

Reference
none (no bf16 baseline; publisher says one cannot run on a single card)
Eval data
wikitext-2, 128 chunks x 2048 (262,016 tokens); AIME-2026 n=30 avg@8
Context
2048
Tool
wikitext2_ppl_vllm; glq-bench aime_2026
  • No bf16 baseline; publisher calls the PPL a sanity check, not a quality claim [L209-211]
  • Calibration is WikiText-2 and PPL is measured on wikitext-2 [L25, L206]
  • AIME-2026 99.58% is at the ceiling with no baseline; publisher says it does not establish quantization cost [L235-237]
  • Card's PPL comparison table is for a different model (Qwen3.8-27B); the implied bf16 ~4.81 is an extrapolation [L213-224]
  • Single runs [L187, L226]

Card snapshot cc48093a06f3 · updated 2026-10-06

yamz-labs/Qwen3.8-Flash-Next-EXL3-Yamz

EXL3 Kyojin 1.5.7+ (ExLlamaV3-based), Strix Halo gfx1151 other or no reference

EXL3 quantization (conversion tooling not published)

Reference
official FP8 release (not BF16)
Eval data
200 public neutral texts, 3084 scored positions, fixed-seed points
Tool
one scorer for all engines, row-cluster bootstrap; rivals patched to emit top-20 log-probabilities
  • Reference is the official FP8 release, not BF16 [L79]
  • Rival rows measure engine + pack together (Strata, Gufo runtimes), patched to output top-20 log-probabilities [L56, L79]
  • Larger Gufo/UD-Q4_K_XL pack has lower KL (0.0156) and higher top-1 than this pack [L85]
  • Pack needs about 113 GiB of the 128 GB in use [L53]
  • Optional runtime uncensor preset ships in the repo, off by default, weights unchanged [L89-95]

Card snapshot 779f63f76b28 · updated 2026-10-16

YTan2000/Qwen3.8-Flash-Next-TQ3_4S

GGUF turbo-tan/llama.cpp-tq3 fork only other or no reference

TurboQuant TQ3_4S (~4 bpw), re-quantized from unsloth's F16 GGUF using unsloth's imatrix (926 entries / 45 chunks); text-only (no vision, no MTP)

Reference
none (no BF16 or other baseline)
Eval data
Hard86 (86 tasks), HumanEval+, MBPP+, in-house quality suite (12–15 items per category)
Context
32768
Tool
official scorers (EvalPlus)
  • No baseline: benchmark scores are absolute, with no BF16 or other-quant comparison [L107-121]
  • Quality-suite categories have 12–15 items each [L117-121]
  • Quality-suite score labels don't match pass counts in places (85.1 for 9/15, 83.3 for 10/15, 81.7 for 12/15) [L118-120]
  • Quantized from unsloth's F16 GGUF with unsloth's imatrix, not from Qwen's BF16 [L135-141]

Card snapshot 589e20f6f865 · updated 2026-09-12

zankich/Qwen3.8-Flash-Next-W4A16-Merlin-INT4PLE-FP8KV

compressed-tensors W4A16 (GPTQ) + INT4 PLE + FP8 KV scales zankich/vllm fork (int4 PLE plugin + qwen4_exp), TP=4 + expert parallel other or no reference

Repack of halt95/Qwen3.8-Flash-Next-W4A16-Merlin (W4A16 GPTQ body, MTP bf16 unchanged) with n-gram PLE table converted FP8 -> INT4 g32 symmetric and calibrated per-layer static FP8 E4M3 KV scales (12 QSA layers, margin 1.20)

Reference
none for the 700-question run (absolute score); FP8-KV subset compared against same model at default KV dtype
Eval data
MMLU-Pro 700 questions (5-shot, chat template); 112-question deterministic subset for FP8 KV
Tool
lm-eval
  • MMLU-Pro on a 700-question sample; FP8-KV comparison on a 112-question subset (±0.031) [L98-103]
  • No BF16 baseline; the omlx 4-bit figure is quoted from a published measurement on a different 300-question subset [L99-100]
  • W4A16 body and MTP head are halt95's checkpoint unchanged; this repo adds INT4 PLE repack and FP8 KV scales [L19-25]
  • 95 GB total; INT4 PLE table 26.8 GB pinned in host RAM [L29-30, L82]

Card snapshot ba8fff329b87 · updated 2026-10-05

Repos without extracted claims (238)

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.

RepositoryStatusFormatWeightsNoteCard
30cmtonyfaker/Qwen3.8-Flash-Next-REAP-288-MLX-Serve-4bit-MTPread in full: no quality claimsMLXmodified weightsNo quality claims; only a 4/4 two-case API smoke test [L33-35]b8a6c8d
AetherKelly00/Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUFNo quality claims for the quant; card reproduces Qwen's base-model tablesb5ae04a
agentionai/Qwen3.8-Flash-Next-MTP-Q8_0-GGUFread in full: no quality claimsGGUF (MTP draft head only)No quality claims for the quant (card read in full): only draft speed and acceptance on Radeon 8060S [L53-58]640aed4
agentionai/Qwen3.8-Flash-Next-MTP-ROCmFP4-FAST-GGUFread in full: no quality claimsGGUF (ROCmFP4, MTP draft head only)No quality claims for the quant (card read in full): only speculative-decoding throughput and draft acceptance on Radeon 8060S, 250 tokens [L56-71]5a2cf56
agnosticeng/Qwen3.8-Flash-Next-4bitread in full: no quality claimsMLX-style (lisa runtime)No quality claims for the quant (card read in full): only quantization format, contents and conversion notes [L23-49]97a3bd1
ahadadada/Ahadadada-Qwen3.8-Flash-Next-FP8read in full: no quality claimsFP8No quality claims for the quant; card reproduces Qwen's base-model tables8a311cf
airagrp/Qwen3.8-Flash-Next-mixed-6-8-ple8-mlx-serveread in full: no quality claimsMLX (mlx-serve pack)No quality claims for the quant (card read in full): '8-bit MTP-expert variant ... no measurable quality ... difference' and '--kv-quant 8 ... no quality loss' are stated without numbers [L87-88, L93]b909fe3
Alexsi98/Huihui-Qwen3.8-Flash-Next-abliterated-GGUFread in full: no quality claimsGGUFmodified weightsNo quality claims for the quant (card read in full): card is a description, a llama-cli command and usage warnings only [L18-54]7711281
amd/Qwen3.8-Flash-Next-EXTENSORread in full: no quality claimsEXTENSOR GGUF (not llama.cpp compatible)No quality claims for the quant (card read in full): only file, size, hash and quantization summary [L23-30]c55e64b
antirez/qwen3.8-flash-next-ggufread in full: no quality claimsGGUFNo quality claimsd600fe1
apetersson/Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUFNo quality claims (memory/speed only)9af845f
ArtomYuan/Qwen3.8-Flash-Next-ROCmFPXread in full: no quality claimsGGUF (ROCmFPX extension types)No quality claims for the quant (card read in full): benchmarks are llama-bench throughput only (pp256/pp2048/tg256) [L84-114]; tier labels such as 'balanced quality' / 'maximum quality' are descriptive, with no measured quality numbers [L40-46, L74-82]e194a61
ashbash/Qwen3.8-Flash-Next-MTP-Drafter-GGUFread in full: no quality claimsGGUF (MTP draft head)No quality claims for the quant (card read in full): draft heads only; no acceptance or quality figuresc82053e
ashuaria/Qwen3.8-Flash-Next-GGUF-FineTuningread in full: no quality claimsGGUFNo quality claims for the quant; card reproduces Qwen's base-model tables147b9cc
auggie246/Swift-1.5-Qwen3.8-Flash-Next-W4A16-FP8PLEread in full: no quality claimsAutoRound W4A16 (auto_gptq) + FP8 PLEmodified weightsNo quality claims for the quant (card read in full): card states there are no speed or quality measurements yet [L33-35]aecff07
auto7x/OrcaRouter-Qwen3.8-Flash-Next-Uncensored-ModelOpt-NVFP4-GGUFread in full: no quality claimsNVFP4 (ModelOpt safetensors)modified weightsNo quality claims for the quant (card read in full): card says runtime use is 'not a claim of new benchmark scores' [L53]3747631
AutomatosX/AX-Qwen3.8-Flash-Next-MLX-AXQ-MXFP4-MTPread in full: no quality claimsMLXNo quality claims for the quant (card read in full): card states quality versus BF16 or uniform baselines is 'Not published; no quality-retention claim' [L166, L27-29]4fdb4a6
AutomatosX/AX-Qwen3.8-Flash-Next-MLX-AXQ-MXFP8-MTPread in full: no quality claimsMLX (AXQuant)No quality claims for the quant (card read in full): card states quality versus BF16 is not published and no quality-retention claim is made [L42-43, L164]48b9aa5
axiomofmind/Qwen3.8-Flash-Next-W4A16-NVFP4read in full: no quality claimsNVFP4 (ModelOpt, W4A16)No quality claims for the quant (card read in full): 29-line card with format description only [L15-29]41ef3b0
Baekpica/Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUF (custom qwen4exp schema, ds4)No quality claims; only structural/byte-level verification [L43-60]5dfe6f0
Baekpica/Qwen3.8-Flash-Next-Mixed-Quant-GGUFread in full: no quality claimsGGUF (custom qwen4exp schema, ds4)No quality claims; card states model quality remains provisional pending validation gates [L31, L88-97]151ad0c
bartowski/Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUFNo measured quality claims; qualitative labels onlydb3a185
Beinsezii/Qwen3.8-Flash-Next-GGUF-HALOread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): 11-line card with only the recipe pointer and a note that it is larger than resident memory [L7-11]e1c7270
BenMonkey11/Huihui-Qwen3.8-Flash-Next-abliterated-v2-GGUFread in full: no quality claimsGGUFmodified weightsNo quality claims for the quant (card read in full): card describes abliteration, Strata/llama.cpp usage and usage warnings only [L30-76]85a485c
BlivionIaG/Qwen3.8-Flash-Next-EXL3-3bpw-3instread in full: no quality claimsEXL3No quality claims for the quant (card read in full): no size, KLD, PPL or benchmark figures; card only says outputs should be representative 'modulo expected quantization loss' [L20]ceaf21e
BlivionIaG/Qwen3.8-Flash-Next-EXL3-3bpw-3inst-uncalibratedread in full: no quality claimsEXL3No quality claims for the quant (card read in full): card labels it uncalibrated WIP, 'not a quality release', outputs not representative [L20]5c7f0eb
boxrunner/Qwen3.8-Flash-Next-IQ3_S-GGUFread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): card states 'No quality benchmark is reported here' [L87]ffd258d
boxrunner/Qwen3.8-Flash-Next-mixC-GGUFread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): card states 'No quality benchmark is reported here' [L100]; only speed figures [L77-98]028efbb
brianbellx/Huihui-Qwen3.8-Flash-Next-abliterated-GGUF-copyread in full: no quality claimsGGUFmodified weightsNo quality claims for the quant (card read in full): card is abliteration description, usage warnings and donation info only [L18-54]0dec8ab
butterf1ying/Qwen3.8-Flash-Next-Uncensored-REAP288-MTP-MLX-4bitread in full: no quality claimsMLX (mlx-serve)modified weightsNo quality claims for the quant (card read in full): smoke test only (loads, generates, MTP active; peak ~40.3 GB on M1 Max); no benchmark or speed claim [L81-90, L175]7fef629
cafonez/Qwen3.8-Flash-Next-HC-Q8read in full: no quality claimsGGUF (ROCmFPX types)No quality claims for the quant (card read in full): only layout, run instructions and a decode speed at 262k [L103-105]986da0f
carsenk/Qwen3.8-Flash-Next-ROGUE-NVFP4read in full: no quality claimsNVFP4 (ModelOpt)modified weightsNo quality claims for the quant (card read in full): only coherent-generation and refusal smoke tests (pass/fail) and throughput [L66-68]dbd1690
cc1244578151/Qwen3.8-Flash-Next-GSQ-RCO-GGUFread in full: no quality claimsGGUFCard is byte-identical to ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF @ ed59f920; its quality numbers are ISTA's claims and are not repeated herec14ca5b
cmh/Qwen3.8-Flash-Next-5.33bpw-PLEQ8_0read in full: no quality claimsGGUFNo quality claims for the quant (card read in full): only size/BPW table, recipe and run configurations [L15-130]3b07e4d
CMSManhattan/JiRackDeltaNet_180bread in full: no quality claimsBF16 safetensors (JiRack port); GGUF plannedNo quality claims for the quant; card reproduces Qwen's base-model tables [L234, L276]82d1140
cyankiwi/Qwen3.8-Flash-Next-AWQ-INT4read in full: no quality claimsAWQNo quality claims for the quant; card reproduces Qwen's base-model tablesd39638a
d9beuD/Qwen3.8-Flash-Next-oQ2-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details and bit-width shares [L20-47]120564d
d9beuD/Qwen3.8-Flash-Next-oQ2.5-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only bit-width shares and size [L27-42]9588892
d9beuD/Qwen3.8-Flash-Next-oQ2.5e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details, bit-width shares / imatrix and sensitivity notes [L38-57]c9fd3c3
d9beuD/Qwen3.8-Flash-Next-oQ2.7-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details and bit-share table [L20-46]23bd805
d9beuD/Qwen3.8-Flash-Next-oQ2.7e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only bit-share table and calibration details [L38-57]77f5e63
d9beuD/Qwen3.8-Flash-Next-oQ2e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details, bit-width shares and imatrix expert coverage (99.97%) [L24-58]7680df8
d9beuD/Qwen3.8-Flash-Next-oQ3-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only bit-width shares and size [L27-42]cf5fd1a
d9beuD/Qwen3.8-Flash-Next-oQ3.5-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details, bit-width shares / imatrix and sensitivity notes [L34-46]c6d197d
d9beuD/Qwen3.8-Flash-Next-oQ3.5e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details, bit shares and imatrix coverage (99.97% experts activated) [L24-57]fc2a7d1
d9beuD/Qwen3.8-Flash-Next-oQ3e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only bit-share table and calibration details [L38-57]24e392e
d9beuD/Qwen3.8-Flash-Next-oQ4-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details and bit-width shares [L20-45]12e38a2
d9beuD/Qwen3.8-Flash-Next-oQ4e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only bit-width shares, imatrix coverage (99.97% experts activated) and size [L31-52]0b047d1
d9beuD/Qwen3.8-Flash-Next-oQ5-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details, bit-width shares / imatrix and sensitivity notes [L34-44]da3149c
d9beuD/Qwen3.8-Flash-Next-oQ5e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details, bit shares and imatrix coverage [L24-55]0eb8d6d
d9beuD/Qwen3.8-Flash-Next-oQ6-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only bit-share table [L34-43]fe8b2e3
d9beuD/Qwen3.8-Flash-Next-oQ6e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details, bit-width shares and imatrix expert coverage [L24-54]fb34e33
d9beuD/Qwen3.8-Flash-Next-oQ8-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only size [L26]24a21ad
d9beuD/Qwen3.8-Flash-Next-oQ8e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details, bit-width shares / imatrix and sensitivity notes [L37-46]3db9c56
ddalcu/Qwen3.8-Flash-Next-MLX-Serve-8bitread in full: no quality claimsMLX (mlx-serve pack)No quality claims for the quant (card read in full): 'closest-to-bf16 pack in the family' and iQ-MLX-4.7bpw 'closer to bf16' are stated without any figures [L22, L31]f61a712
densohax/Qwen3.8-Flash-Next-heretic2-W4A16-FP8PLEread in full: no quality claimsAutoRound W4A16 (GPTQ packing) + FP8 PLEmodified weightsNo quality claims for the quant (card read in full): only throughput figures, measured on the upstream albucino checkpoint, not this one [L72-75, L133-134]705a2ce
devan-carlin/Qwen3.8-Flash-Next-W4A16read in full: no quality claimscompressed-tensors W4A16No quality claims (decode speed only) [L40-43]40b8f18
DevelopingDad/Qwen3.8-Flash-Next-Uncensored-NVFP4-DGX-Sparkread in full: no quality claimsNVFP4modified weightsNo quality claims for the quant: the 20-case assistant-reviewed diagnostic (14/20 vs 13/20) compares runtime profiles of the same weights, and publisher says it is not evidence of model quality [L110-112]6773ecf
DevQuasar/Qwen.Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): card is a logo, a slogan and a link to the base model [L7-13]ef7dd83
DhruvalLabs/Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): 'Quality' column is a star rating (4 stars) with no measurement behind it [L41-45]c257284
doth4580/Qwen3.8-Flash-Next-EXL3-3.05bpwread in full: no quality claimsEXL3No quality claims for the quant (card read in full): only throughput/latency figures on DGX Spark; quality is left to the shipped qbench prompts for users to reproduce [L111-147, L191-192]2f41a00
doth4580/Qwen3.8-Flash-Next-EXL3-4.05bpwread in full: no quality claimsEXL3Quality check is a next-token-loss comparison with the publisher's own 3.05 pack (0.625 vs 0.621 at 8K, 1.579 vs 1.578 at 16K), which the card calls a mis-decode alarm, not a quality ranking [L152-154]e2e53fd
drluoto/Qwen3.8-Flash-Next-MTP-GGUFread in full: no quality claimsGGUF (MTP draft head only)No quality claims for the quant (card read in full): draft head only; all figures are whole-stack decode throughput and draft acceptance, not quantization quality [L76-98, L118-124]ac875a9
drowzeys/keys-Qwen3.8-flash-next-ablit-Mia-Single-Spark-onlyread in full: no quality claimsnot statedmodified weightsNo quality claims for the quant (card read in full): README snapshot is emptyb3b60a4
Dyluhn/Qwen3.8-Flash-Next-R9V-IQ4_XSread in full: no quality claimsGGUF (bundle)No quality claims for the quant (card read in full): only provenance and a link to an external performance qualification report [L51-56]f48e978
dzannotti/Qwen3.8-Flash-Next-MTP-GGUFread in full: no quality claimsGGUF (MTP draft head)No quality claims for the quant (card read in full): draft head only; 'a Q8_0 head measured worse' refers to draft agreement, and tables report speed/acceptance [L35, L91-97]0b2551d
EasiiX/Qwen3.8-Flash-Next-MTP-Strix-Halo-GGUFread in full: no quality claimsGGUF (MTP sidecar only)No quality claims for the quant (card read in full): only speculative-decoding speed/acceptance figures on Strix Halo (decode +50% on code, ~85% acceptance) [L23-26]896756c
edougawa/Qwen3.8-Flash-Next-W4A16-PLE4-MTP-Sparkread in full: no quality claimscompressed-tensors W4A16No quality claims; only structural validation PASS, all runtime/quality checks 'Not tested' [L142-151]9f10435
eemin/halogen-qwen3.8-flash-next-copyread in full: no quality claimshalogen .hgnNo quantitative quality claims vs BF16; only qualitative statements ('roughly 5–9% worse on perplexity' without the sidecar, o_proj at 8-bit 'a statistical tie with full BF16 rows') [L57-59, L124-130]c01ee53
EigenLabs/Qwen3.8-Flash-Next-MLX-4bit-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only compatibility pass counts (59 standard cases, 24 tool-history cases) and integrity checks; card states these do not establish BF16 equivalence [L102-107, L115]e555958
fabiogreter/Qwen3.8-Flash-Next-lily-q4read in full: no quality claimslily qwen4_exp-affine-v1 (safetensors, engine-specific)No quality claims for the quant (card read in full): only speed figures; card notes quantized output is not bit-identical to bf16 [L117-137, L145-146]28ed324
FalconZypp/Qwen3.8-Flash-Next-AWQ-INT4read in full: no quality claimsAWQNo quality claims for the quant; card reproduces Qwen's base-model tables [L112-134]2b1c486
gabmiranda/Qwen3.8-Flash-Next-oQ2.5e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): card is YAML front matter and a title only [L1-11]e372f69
gabmiranda/Qwen3.8-Flash-Next-oQ2.7e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): card is front matter plus a title only [L1-11]79f0ee1
gabmiranda/Qwen3.8-Flash-Next-oQ2e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): card is only front matter and a title [L1-11]9a235a2
gabmiranda/Qwen3.8-Flash-Next-oQ3.5e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): card is only front matter and a title [L1-11]cf29543
gabmiranda/Qwen3.8-Flash-Next-oQ3e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): 11-line card with front matter and a title only [L1-11]d1453d3
gabmiranda/Qwen3.8-Flash-Next-oQ4e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): card is YAML front matter and a title only [L1-11]e3ce05a
gabmiranda/Qwen3.8-Flash-Next-oQ5e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): card is front matter plus a title only [L1-11]cf2679f
gabmiranda/Qwen3.8-Flash-Next-oQ6e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details (bits, group size) [L13-20]0d4741f
gabmiranda/Qwen3.8-Flash-Next-oQ8e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details [L13-20]7453cc0
garnermccloud/Qwen3.8-Flash-Next-MLX-SSD-Streamread in full: no quality claimsMLX (mlx-serve pack)No quality claims for the quant (card read in full): 'Quality-first' heading but only speed figures, one long-context retrieval pass and 39 pass/fail end-to-end checks with no scores [L21-26, L66-75]9a624d9
GBP-DE/Qwen3.8-Flash-Next-oQ4e-WLSRefit-MTPread in full: no quality claimsMLX (oMLX oQ)No quality claims for the quant (card read in full): only qualitative observations from a GPT-6-reviewed 60-question 'Silbergrat Reasoning Challenge' and a trading-research prompt, which the card itself calls 'qualitative, not standardized benchmarks' [L35-62]3f3eb3a
GBP-DE/Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTPread in full: no quality claimsMLX (oMLX oQ5e)No measured quality claims; card says practical quality 'has not yet been independently benchmarked' [L47-50]559775f
GCSA-AiLab/Qwen3.8-Flash-Next-FP8-Abliterated-GGUFread in full: no quality claimsGGUFmodified weightsNo quality claims for the quant (card read in full): card is only links to per-language docs and download folders [L36-42]881b6c5
ggml-org/Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUFNo quality claims; card has a 'TODOs: add info' placeholder [L29-31]0a61d2b
groxaxo/Qwen3.8-Flash-Next-exl3-3.50bpw_hq_h6_ng6read in full: no quality claimsEXL3No quality claims for the quant (card read in full): only load ceilings, throughput and MTP acceptance; card states these are not a quality score vs BF16 [L86-169]5ebeaa0
Guile/Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): only file sizes and qualitative per-file descriptions ('Good quality', 'Very low quality but surprisingly usable') with no measured numbers [L42-71]75862fa
Hagwell/Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUFNo quality claims for the quant; card reproduces Qwen's base-model tables339aed9
haihengh/Qwen3.8-Flash-Next-125B-finch-4bitread in full: no quality claimsFinchMoE .finch (custom streaming layout)No quality claims for the quant (card read in full): only an oracle argmax-agreement correctness check; card states no downstream benchmark has been run and none is claimed [L164-169]2464910
him0413/Qwen3.8-Flash-Next-Uncensored-Q4_K_M-MTPread in full: no quality claimsGGUFmodified weightsNo quality claims for the quant (card read in full): only model details and usage [L37-67]54fc513
hstolte/Qwen3.8-Flash-Next-HALOLOOM-INT4-INT8-AWQread in full: no quality claimsQuark mixed INT4/INT8 (AWQ)No quality claims for the quant (card read in full): card states independent model-quality evaluation is not claimed [L215]dadedd9
htdohk/Huihui-Qwen3.8-Flash-Next-abliterated-MTP-GGUFread in full: no quality claimsGGUFmodified weightsNo quality claims for the quant (card read in full): description, source notes, a llama-cli command and usage warnings only [L27-82]83eeb1c
huihui-ai/Huihui-Qwen3.8-Flash-Next-abliterated-GGUFread in full: no quality claimsGGUFmodified weightsNo quality claims for the quant (card read in full): card is abliteration description, update notes, usage warnings and donation info only [L27-88]06a318d
huihui-ai/Huihui-Qwen3.8-Flash-Next-abliterated-v2-GGUFread in full: no quality claimsGGUFmodified weightsNo quality claims for the quant (card read in full): card describes abliteration, Strata/llama.cpp usage and usage warnings only [L30-76]43b4c1a
igorls/Qwen3.8-Flash-Next-mixed-NInferread in full: no quality claimsNInfer container (NVFP4/FP8 + INT4 PLE)No quality claims for the quant (card read in full): card states it 'adds no benchmark campaign or quality score' [L207-208]ef61b3c
ilintar/qwen3.8-flash-next-gguf-strix-haloread in full: no quality claimsGGUFNo quality numbers; only states quality 'was checked by paired perplexity against a reference stack' [L92-93]ba5b0d6
immiq/Qwen3.8-Flash-Next-Orinfer-E8P-Q2A8read in full: no quality claimsOrinfer safetensors (E8P)No quality claims for the quant (card read in full): card states a full LLM quality benchmark remains to be added; only prefill/decode speed [L40, L58-80]4c1ffc2
Inferact/Qwen3.8-Flash-Next-NVFP4read in full: no quality claimsNVFP4No quality claims for the quant; card reproduces Qwen's base-model tables103a760
inferencerlabs/Qwen3.8-Flash-Next-MLX-Q4iread in full: no quality claimsMLXNo quality claims for this model's quant: card states the quality table is from the conversion of GLM 5.1, not Qwen3.8-Flash-Next [L27]85303bc
jlkivey/Qwen3.8-Flash-Next-MTP-PR27836-GGUFread in full: no quality claimsGGUFNot a quant of the main model: draft-head export only; measurements are decode speed and draft acceptance on unsloth UD-IQ4_XS [L56-68]159500b
jmgao/Swift1.5-Flash-Next-mixed-NInferread in full: no quality claimsNInfer mixed (NVFP4/FP8/INT4)modified weightsNo quality claims for the quant (card read in full): three-paragraph description only [L28-34]2f012c4
Joakimpalm-Zen/Qwen3.8-Flash-Next-MTP-GGUFread in full: no quality claimsGGUF (MTP draft head)No quality claims for the quant (card read in full): this is a draft head; card reports only decode speed and draft acceptance (69-82%) [L35-47, L92-94]c104e65
jockevaupptaget/Qwen3.8-Flash-Next-MTP-GGUFread in full: no quality claimsGGUF (MTP draft head only)No quality claims for the quant (card read in full): draft head only; figures are decode speedup and draft acceptance with a UD-Q3_K_XL target [L24-25]69da733
jpezzulli/OrcaRouter-Qwen3.8-Flash-Next-Uncensored-ModelOpt-NVFP4read in full: no quality claimsNVFP4 (ModelOpt 0.46.0)modified weightsNo quality claims for the quant (card read in full): only daily-use runtime evidence; card explicitly makes no benchmark claim [L45]f24d2b6
Jrouth/Qwen3.8-Flash-Next-MTP-C2T8-refit-GGUFread in full: no quality claimsGGUF (MTP draft head only)modified weightsNo quality claims for the quant (card read in full): only draft acceptance and decode speed; card states the target re-checks every token so output is unchanged [L17-18, L79-91]7876390
junafinity/Qwen-3.8-Flash-Next-Uncensored-MLX-MXFP4read in full: no quality claimsMLXmodified weightsNo quality claims for the quant: card states the file has not been scored with PPL, KLD or evals [L69, L147, L239]6fd6569
Jundot/Qwen3.8-Flash-Next-oQ2.5e-fp16-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only bit distribution and storage tables [L31-53]9eed92c
Jundot/Qwen3.8-Flash-Next-oQ2.5e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only bit-distribution and storage tables [L28-53]ebfb104
Jundot/Qwen3.8-Flash-Next-oQ2.7e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only bit distribution and storage breakdown [L29-50]0e751a5
Jundot/Qwen3.8-Flash-Next-oQ3.5e-fp16-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only bit distribution and storage tables [L31-55]42ce563
Jundot/Qwen3.8-Flash-Next-oQ3.5e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only bit distribution and storage tables [L32-55]e821605
Jundot/Qwen3.8-Flash-Next-oQ3e-fp16-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only bit distribution and storage tables [L31-55]14f743e
Jundot/Qwen3.8-Flash-Next-oQ3e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only bit-distribution and storage tables [L28-55]b73babc
Jundot/Qwen3.8-Flash-Next-oQ4e-fp16-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only bit distribution and storage breakdown [L31-54]3ed70e9
Jundot/Qwen3.8-Flash-Next-oQ6e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only bit distribution and storage tables [L29-48]816d1eb
Jundot/Qwen3.8-Flash-Next-oQ8e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only bit distribution and storage tables [L29-46]e6658a3
KaedeTai/Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp-zhTWread in full: no quality claimsMLX (per repo name)modified weightsNo quality claims for the quant (card read in full): README snapshot at this sha is empty (0 bytes)c4d04ba
Kandandan/Qwen3.8-Flash-Next-MLX-4bit-MTPread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only speculative-decoding acceptance rates and speedups (M3 Max, 160 tokens x 10 tasks) [L48-60]40e15b0
KENIC-1/Qwen3.8-Flash-Next-GGUF-UD-IQ4_XSread in full: no quality claimsGGUFNo quality claims for the quant; card reproduces Qwen's base-model tablesc639d07
KENIC-1/Qwen3.8-Flash-Next-GGUF-UD-Q4_K_XLread in full: no quality claimsGGUFNo quality claims for the quant; card reproduces Qwen's base-model tables6bdfee4
kingjones777/Qwen3.8-Flash-Next-MTP-Heads-GGUFread in full: no quality claimsGGUF (MTP draft heads)No quality claims for the quant (card read in full): draft heads; only speed and acceptance figures [L39-48]72805b6
kingjones777/Qwen3.8-Flash-Next-ROCmFP4-FAST-GGUFread in full: no quality claimsGGUF (ROCmFP4)No quality claims for the quant (card read in full): card states 'Quality is the one thing I have not measured, so I won't claim it' [L297-298]7565b1e
kingjones777/Qwen3.8-Flash-Next-ROCmFP4-STRIX-GGUFread in full: no quality claimsGGUF (ROCmFP4, Q4_0_ROCMFP4_STRIX)No quality claims for the quant (card read in full): card states 'Quality is the one thing I have not measured, so I won't claim it' [L296-298]7320914
kingjones777/Qwen3.8-Flash-Next-ROCmFP4-STRIX_LEAN-GGUFread in full: no quality claimsGGUF (ROCmFP4)No quality claims for the quant (card read in full): card explicitly states 'Quality is the one thing I have not measured, so I won't claim it' [L297-298]bf7fbc6
Litwein/Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MTPLXread in full: no quality claimsMLX (MTPLX layout)modified weightsNo quality claims for the quant (card read in full): only speed/MTP A/B and memory figures; card says no task or vision-benchmark score is claimed and refers to the MLX repo for teacher-gap tables [L59, L91-95, L104-115, L152-153]2f7b82f
lmstudio-community/Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUFNo quality or size claims on card; sizes from HF listing158fc82
lovedheart/Qwen3.8-Flash-Next-NVFP4-W4A16-ATTN-FP8-MTP-NVFP4read in full: no quality claimsNVFP4 (ModelOpt mixed: NVFP4 W4A16 experts + FP8 block attention)No quality claims for the quant (card read in full): only MTP draft acceptance figures (accept_length 2.5–2.6, accept_rate ~0.53) [L73-81]f748827
Lucebox/Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): only Lucebox decode/prefill speed figures [L32-39]f048355
LunarMoonar/Huihui-Qwen3.8-Flash-Next-abliterated-GGUFread in full: no quality claimsGGUFmodified weightsNo quality claims for the quant (card read in full): description, source notes, a llama-cli command and usage warnings only [L27-88]e0db1ef
maclocal/Qwen3.8-Flash-Next-AFM-MLX-4bitread in full: no quality claimsMLX (AFM)No quality claims for the quant; card reproduces Qwen's base-model tables [L117, L181-203]2ac1a93
madlabs-ai/Qwen3.8-Flash-Next-NVFP4-MTP-BF16read in full: no quality claimsNVFP4 (ModelOpt)No quality claims for the quant (card read in full): 'gets NVIDIA's quantization quality' is stated without figures; only decode speed and a tool-call malformation count tied to an SGLang strictness setting [L65-66, L85-96]85bee1e
mazinb/Qwen3.8-Flash-Next-Uncensored-NVFP4read in full: no quality claimsNVFP4modified weightsNo quality claims for the quant (card read in full): the README snapshot at this sha is empty (0 bytes)f2c21eb
mduca771/Qwen3.8-Flash-Next-NG8-GGUFread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): states PLE compression has 'direct and perceptible' quality impact and NG8 gives near-Q8_0 embedding quality, without measurements [L25, L33]; validation is tensor/byte integrity plus smoke test [L39-44]c516e96
meshapplied/Qwen3.8-Flash-Next-NVFP4-Latestread in full: no quality claimsNVFP4 (ModelOpt, per HF tags)No quality claims for the quant (card read in full): README snapshot at this sha is empty (0 bytes); format/runtime from HF tags only13b21e1
MESHIVEAI/Qwen3.8-Flash-Next-NVFP4-W4A16read in full: no quality claimsNVFP4 (ModelOpt, W4A16)No quality claims for the quant (card read in full): card has no evaluation and states serving has not been validated end to end [L68-70]d58bb7a
mkinc/Qwen3.8-Flash-Next-rcO4-GGUFread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): only storage-by-group table and Strata packing instructions [L23-98]b98c43c
mlx-community/Qwen3.8-Flash-Next-4bitread in full: no quality claimsMLXNo quality metrics (KLD/PPL/benchmarks); 'Verification' table checks conversion correctness only [L37-45]07b5dc6
mlx-community/Qwen3.8-Flash-Next-4bit-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): 27-line card describing contents only [L13-27]825df62
mlx-community/Qwen3.8-Flash-Next-OptiQ-2bitread in full: no quality claimsMLX (mlx-optiq)No quality claims for the quant (card read in full): card states no Capability Score is published; only a Flappy Bird demo [L31-35, L49]52a21f4
mlx-community/Qwen3.8-Flash-Next-oQ6e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): benchmark tables are speed only (write/read tok/s, tokens written) vs oQ4e and oQ8e [L28-62]e171af8
mlx-community/Qwen3.8-Flash-Next-oQ8e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only speed benchmark vs Jundot oQ4e and an oQ6e pack (time to completion, tok/s) [L28-64]2a9de02
MonumentalSystems/Qwen3.8-Flash-Next-control-vectorsread in full: no quality claimsGGUF control vectorsNo quality claims for the quant (card read in full): repo contains activation-steering vectors, not model weights; measured tables report steering effects (token-length deltas, comment ratios), not quantization fidelity [L99-110, L141-155]455a5a1
mradermacher/Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): only sizes and generic qualitative notes ('lower quality', 'fast, recommended', 'very good quality', 'fast, best quality') [L40-54]68d514c
mradermacher/Qwen3.8-Flash-Next-i1-GGUFread in full: no quality claimsGGUFNo measured quality claims; notes are generic boilerplate labels, not measurements of this model [L45-58]9547b1c
msuiche/Qwen3.8-Flash-Next-abliterated-cyber-GLP-47read in full: no quality claimsnot statedmodified weightsNo quality claims for the quant (card read in full): README snapshot at this sha is empty (0 bytes)bb05ede
myllmbox/Qwen3.8-Flash-Next-hibrid47-uncensoredread in full: no quality claimscompressed-tensors NVFP4/FP8 (per HF tags)modified weightsNo quality claims for the quant (card read in full): README snapshot at this sha is empty (0 bytes); format from HF tags onlye2edbc4
myllmbox/Qwen3.8-Flash-Next-hibrid48-uncensoredread in full: no quality claimsnot statedmodified weightsNo quality claims for the quant (card read in full): the README snapshot at this sha is empty (0 bytes)572e216
Nielk38/Qwen3.8-Flash-Next-GGUF-SPLITread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): only a 2+2 smoke test and tensor checksum match [L96-98]b816449
nitinpanj/Qwen3.8-Flash-Next-Q4_0-Q8out-v3-GGUFread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): card says it was 'built and benchmarked locally' but gives no numbers [L61-62]967e141
nmatteo3294/Qwen3.8-Flash-Next-Q3_PLE-MTP-GGUFread in full: no quality claimsGGUF (private Q3_PLE type, patched llama.cpp)No quality claims for the quant (card read in full): 'exact output'/parity checks compare MTP vs target-only decoding, not quantization quality [L105-111, L138-141]6650250
nopmobiel/Qwen3.8-Flash-Next-MLX-4bitread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only throughput and peak memory; functional chat/image checks without scores [L76-91]8577fa9
orcarouter/Qwen3.8-Flash-Next-Uncensored-FP8read in full: no quality claimsFP8 (per repo name)modified weightsNo quality claims for the quant (card read in full): README snapshot at this sha is empty (0 bytes)a920884
orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUFread in full: no quality claimsGGUFmodified weightsNo quality claims for the quant (card read in full): README snapshot is emptyd27c0a1
orcarouter/Qwen3.8-Flash-Next-Uncensored-MLXread in full: no quality claimsMLX (per repo name)modified weightsNo quality claims for the quant (card read in full): README snapshot at this sha is empty (0 bytes)fb8c6ba
orcarouter/Qwen3.8-Flash-Next-Uncensored-NVFP4read in full: no quality claimscompressed-tensors NVFP4/FP8 (per HF tags)modified weightsNo quality claims for the quant (card read in full): README snapshot at this sha is empty (0 bytes); format from HF tags onlycd56f6c
p4ik/Qwen3.8-Flash-Next-MLX-BEAM-4bitread in full: no quality claimsMLX (B.E.A.M. layout)No quality claims for the quant (card read in full): card states quality was not measured by the publisher and refers to the source build's card [L21, L52, L81]29e259b
paperniuk/Qwen3.8-Flash-Next-MTP-blockread in full: no quality claimsGGUF (MTP block only)No quality claims for the quant (card read in full): only speculative-decoding speeds and draft acceptance; output identical to plain decoding [L52-64]d98c24a
peasantsmith/Qwen3.8-Flash-Next-PS-IQ2_XXS-GGUFread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): only single-prompt pass/fail checks (tool call, needle, smoke) [L45-53]; perplexity explicitly not measured [L54-56]c8122ff
peonist-ai/halogen-qwen3.8-flash-nextread in full: no quality claimshalogen .hgn (custom)No numeric quality claims for the quant (card read in full): only qualitative statements — 8-bit o_proj 'measures as a statistical tie with full BF16 rows', sidecar holds 'most of the quality' — with no figures [L174-180]43898d3
Pragmatism0220/Huihui-v2-Qwen3.8-Flash-Next-GSQ-RCO-Strata-abliterated-direction-GGUFread in full: no quality claimsGGUF control vectormodified weightsNo quality claims for the quant (card read in full): repo is a control vector, not a quantized model; all figures are weight-space extraction/validation metrics (rank-one energy 99.833 %, cos 0.99999998 vs third-party extraction, cos 0.9999967 vs v2) [L163-191]8003251
pugant/Qwen3.8-Flash-Next-MTP-DRAFT-HEAD-ROCMFP4-GGUFread in full: no quality claimsGGUF (MTP draft head only; ROCmFP4 output tensor)No quality claims for the quant (card read in full): only draft decode speed and acceptance; card says greedy output is bit-identical in a 12-prompt battery [L51-58]ba3fe94
Quat3rnion/halogen-swift1.5-qwen3.8-flash-next-v2read in full: no quality claimsHGN (Halogen)modified weightsNo quality claims for current files; card states they have not been loaded or inference-qualified and historical results do not apply [L34, L55, L69]d6faf17
Quat3rnion/Signal-3.8-Flash-Next-HGN-v2read in full: no quality claimsHGN (Halogen)modified weightsNo quality claims for the quant (card read in full): section titled 'performance, not quality'; card states no IFEval or BFCL evaluation completed and no claim about accuracy, retained quality or BF16 equivalence [L154, L188-190]d93045e
quimmedes/Qwen3.8-Flash-Next-MTP-GGUFread in full: no quality claimsGGUF (MTP draft head only)No quality claims for the quant (card read in full): draft heads only; only an approximate acceptance-rate range (~50-70%) is mentioned [L78]e56ac7b
Qwen/Qwen3.8-Flash-Next-FP8read in full: no quality claimsFP8Official; no FP8-specific quality numbers on card236dfdf
RESMP-DEV/Qwen3.8-Flash-Next-IQ2_XS-Q2_0-Standard-iMatrix-GGUFread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): card states no perplexity, task-quality, long-generation or broad throughput evaluation has been run; checks prove integrity and execution only [L31-32, L133-136]694d57e
risk-magician/Qwen3.8-Flash-Next-AMX_Q8R.amxqread in full: no quality claimsAMX_Q8R (.amxq)No quality claims for the quant; card reproduces Qwen's base-model tables50647bc
robertbschuette/Huihui-Qwen3.8-Flash-Next-abliterated-GGUFread in full: no quality claimsGGUFmodified weightsNo quality claims for the quant (card read in full): no sizes, KLD, PPL or benchmarks24b9683
RockTalk/Qwen3.8-Flash-Next-MLX-4bitread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only throughput, MTP acceptance and qualitative vision/tool tests [L66-73, L107-134]0afead4
SargeDev/qwen3.8-flash-next-lil-rascal-uncensoredread in full: no quality claimsNVFP4/MXFP8modified weightsNo quality claims for the quant (card read in full): card is provenance, intended/prohibited use and disclaimers only [L23-85]e20e53c
Sawfwair/Qwen3.8-Flash-Next-MLX-4bitread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only conversion inventory and a statement that packing/layout are validated by a bundled JSON [L20-34]6cc9bbc
Sawfwair/Qwen3.8-Flash-Next-MLX-Mixed-2bitread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only conversion validation (tensor inventory/hashes) and run instructions [L140-145]a6e3d7a
scottlowry/Qwen3.8-Flash-Next-oQ3.5e-fp16-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details [L13-20]d203e0a
scottlowry/Qwen3.8-Flash-Next-oQ3.5e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details [L13-20]2c1f59f
scottlowry/Qwen3.8-Flash-Next-oQ4e-fp16-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): 20-line card with quantization parameters only [L11-20]aa8cb7c
scottlowry/Qwen3.8-Flash-Next-oQ4e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only a four-line quantization-details list [L15-20]998d2d5
scottlowry/Qwen3.8-Flash-Next-oQ5e-fp16-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details [L13-20]f876fb2
scottlowry/Qwen3.8-Flash-Next-oQ5e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details [L13-20]38b03ce
scottlowry/Qwen3.8-Flash-Next-oQ6e-fp16-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details [L13-20]0d81135
scottlowry/Qwen3.8-Flash-Next-oQ6e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): 20-line card with quantization parameters only [L11-20]9cb47a7
scottlowry/Qwen3.8-Flash-Next-oQ8e-fp16-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only a four-line quantization-details list [L15-20]6e1875b
scottlowry/Qwen3.8-Flash-Next-oQ8e-mtpread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only quantization details [L13-20]4bf7525
scouzi1966/Qwen3.8-Flash-Next-AFM-MLX-4bitread in full: no quality claimsMLX (AFM)No quality claims for the quant; only a decode-speed smoke test on a numbered-list prompt [L88-105]; card reproduces Qwen's base-model tables [L117+]3114e33
spiritfather/cyberneurova-Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUFmodified weightsNo quality claims for the quant (card read in full): 'lower quality' / 'very good quality' notes are generic labels without measurements [L40-47]; CaliperBench scores are linked, not shown [L31-32]cd72aeb
starkweatherdigital/qwen3.8-flash-next-nvfp4read in full: no quality claimsNVFP4 (ModelOpt)No quality claims for the quant (card read in full): only throughput and MTP acceptance; card points to GitHub for verification/benchmarks [L27-29]1b304e5
stgallenquants/Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUFNo quality claims for the quant; card reproduces Qwen's base-model tables [L89-104]34cf440
SUNNYROAM/Huihui-Qwen3.8-Flash-Next-abliterated-GGUFread in full: no quality claimsGGUFmodified weightsNo quality claims for the quant (card read in full): description, source notes, a llama-cli command and usage warnings only [L27-88]3110adf
SZYBKIMARCIN955/Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUFNo quality claims for the quant; card reproduces Qwen's base-model tables5ef30cd
talhasarac/Qwen3.8-Flash-Next-REAP128-86B-Unity-OpenCode-GGUFread in full: no quality claimsGGUFmodified weightsNo quality claims for the quant (card read in full): card states no broad benchmark suite has been run; validation notes are anecdotal and report clear general-quality degradation [L42, L143-147]d396d76
tapsin/TAPSIN-Qwen3.8-4B-Distill-GGUFread in full: no quality claimsGGUFmodified weightsNot a quant of Qwen3.8-Flash-Next: a distillation into the Qwen3.5-4B architecture despite base_model: Qwen/Qwen3.8-Flash-Next [L7, L14, L68]994c589
tarruda/Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): only recipe table and a launch script [L22-50]0eb873e
tawei-amd/Qwen3.8-Flash-Next-strixflash-GGUFread in full: no quality claimsGGUFNo quality numbers: Accuracy section says perplexity 'measurement pending' [L70]2cdd89a
tcclaviger/Qwen3.8-Flash-Next-MXFP4-FP8read in full: no quality claimsMXFP4/FP8No quality claims for the quant; card reproduces Qwen's base-model tables58ccaf6
TensorFold/Qwen3.8-Flash-Next-MLX-2bit-MTPread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only throughput, MTP acceptance and pass/fail smoke gates (instruction, factual, arithmetic, coherence) with no scores [L111-120, L153]28b1090
TensorFold/Qwen3.8-Flash-Next-MLX-3bit-MTPread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only throughput, MTP acceptance and pass/fail parity/coherence checks with no scores [L117-126, L175-182]1d53b5c
TensorFold/Qwen3.8-Flash-Next-MLX-4bitread in full: no quality claimsMLXNo quality claims for the quant; only decode-speed figures and a smoke test [L100-113, L142-143]4fc403b
TensorFold/Qwen3.8-Flash-Next-MLX-4bit-MTPread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only MTP speed/acceptance and pass/fail parity/coherence checks with no scores [L117-126, L175-182]2b170fa
TensorFold/Qwen3.8-Flash-Next-MLX-6bit-MTPread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only MTP on/off throughput, draft acceptance and pass/fail parity/coherence checks [L114-129, L172-182]45032a1
TensorFold/Qwen3.8-Flash-Next-MLX-8bit-MTPread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only MTP speed/acceptance and pass/fail checks (MTP on/off exact-output parity, coherence) [L120-126, L178-182]a98d2ff
TensorFold/Qwen3.8-Flash-Next-MLX-oQ2read in full: no quality claimsMLXNo quality claims; card carries a quality-hold warning that live testing found incoherent output [L53]75e9a28
TensorFold/Qwen3.8-Flash-Next-MLX-oQ2-MTPread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only throughput, MTP acceptance and pass/fail smoke checks with no scores [L113-122, L151-156]0a1bf6f
TensorFold/Qwen3.8-Flash-Next-MLX-oQ3-MTPread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only throughput, MTP acceptance and pass/fail exact-answer/coherence gates with no scores [L109-118, L147-151]993186e
TensorFold/Qwen3.8-Flash-Next-MLX-oQ4read in full: no quality claimsMLXNo quality claims (speed-only measurements and generation smoke tests) [L101-115, L141-144]dde9d8d
TensorFold/Qwen3.8-Flash-Next-MLX-oQ4-MTPread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only short-prompt speed tests (27-71 output tokens) and qualitative coherence notes [L111-119, L149-153]069090c
TensorFold/Qwen3.8-Flash-Next-MLX-oQ6read in full: no quality claimsMLXNo quality claims (speed-only measurements and generation smoke tests)aabd8cd
TensorFold/Qwen3.8-Flash-Next-MLX-oQ6-MTPread in full: no quality claimsMLX (oQ)No quality claims for the quant (card read in full): only short-run speeds, MTP acceptance and coherence checks [L104-123, L147-154]d6209ba
TensorFold/Qwen3.8-Flash-Next-MLX-oQ8-MTPread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only MTP speed and deterministic smoke gates (hello / Paris / arithmetic) [L110-116]eb86370
thoddnn/Qwen3.8-Flash-Next-4bitread in full: no quality claimsMLXNo quality claims for the quant (card read in full): verification covers norm-gain conversion correctness and coherent generation only [L37-45]b070255
thousandml/Qwen3.8-Flash-Next-CQ-4bitread in full: no quality claimsCQ 4-bit (custom)No quality claims for the quant (card read in full): only qualitative statements that CQ is 'designed to preserve' behavior / generalize, with no figures [L27, L35, L176]a4492d6
tirex2001/Qwen3.8-Flash-Next-DACANread in full: no quality claimsDACAN pack (GGUF dense + expert .bin)No quality claims for the quant (card read in full): only speed figures and a check that three greedy answers match the service's file character for character [L22-23, L77-88]bbd468e
ToPo-ToPo/Qwen3.8-Flash-Next-mlx-4bitread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only conversion command and usage [L14-52]f128069
ToPo-ToPo/Qwen3.8-Flash-Next-mlx-8bitread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only conversion command and usage [L14-52]ac6a8c8
Trnio/qwen-3.8read in full: no quality claimsFP8No quality claims for the quant; card reproduces Qwen's base-model tablesd44e7a0
trustos/Qwen3.8-Flash-Next-GSQ-RCO-Q2-MTPLXread in full: no quality claimsMLX (MTPLX)No quality claims for the quant (card read in full): Quality section says 'Not evaluated'; ISTA's evaluation covers a different dense-layer pairing [L147-155]09d7bd1
trymirai/Qwen3.8-Flash-Next-S-experimentalread in full: no quality claimsnot statedNo quality claims for the quant (card read in full): README snapshot at this sha is empty (0 bytes)af20c48
trymirai/Qwen3.8-Flash-Next-S-small-experimentalread in full: no quality claimssafetensors (2-bit, per HF tags)No quality claims for the quant (card read in full): README snapshot at this sha is empty (0 bytes); format from HF tags onlyc26c47d
txgsync/Qwen3.8-Flash-Next-MXFP4-MLXread in full: no quality claimsMLXNo quality claims for the quant; card says no comparison with BF16 logits was done and only load/smoke checks were run [L113-125]016f76a
UltimateIntent/Qwen3.8-Flash-Next-Abliterated-s1.5-GGUFread in full: no quality claimsGGUFmodified weightsNo quality claims for the quant (card read in full): only a qualitative statement that evaluation showed 'no noticeable performance / capability loss', with no numbers, method or reference [L76-77]dc11a16
unsloth/Qwen3.8-Flash-Next-FP8read in full: no quality claimsFP8No quality claims for the quant; card reproduces Qwen's base-model tables41cc25f
V1ta11y1/Huihui-Qwen3.8-Flash-Next-abliterated-v2-GGUFread in full: no quality claimsGGUFmodified weightsNo quality claims for the quant (card read in full): card describes abliteration, Strata/llama.cpp usage and usage warnings only [L30-76]09de5fa
vaultai/Qwen3.8-Flash-Next-MLX-oQ4-MTPread in full: no quality claimsMLXNo quality claims for the quant (card read in full): only short-prompt speed tests and qualitative coherence notes [L106-114, L144-148]6b96f4c
vcruz305/Qwen3.8-Flash-Next-EXL3-SAGE-4.15bpwread in full: no quality claimsEXL3No quality claims for the quant (card read in full): card states quality gates have not run and the pack carries no quality claim [L25, L52-54, L72]d5d792a
vcruz305/Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): only approximate backbone sizes; shared BF16 PLE ~95.4 GiB on top [L34-43]0c50dde
Vitruvi0/qwen3.8-flash-next-nvfp4-fp8hybridread in full: no quality claimsNVFP4/FP8 (per repo name)No quality claims for the quant (card read in full): README snapshot at this sha is empty (0 bytes)9a463b4
vkshdev/Qwen3.8-Flash-Next-FP8-180Bread in full: no quality claimsFP8No quality claims for the quant; card reproduces Qwen's base-model tables1e0c4cf
vumpt/Qwen3.8-Flash-Next-GGUFread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): only layout, file and usage [L21-47]83a79fc
Wladastic/Qwen3.8-Flash-Next-Ombro-v13b-GGUFread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): only size/BPW and per-tensor type comparison against UD-Q4_K_XL, Ombro V9 and UD-Q5_K_XL [L23-59]4b20669
Wladastic/Qwen3.8-Flash-Next-Ombro-v9-GGUFread in full: no quality claimsGGUFNo quality claims for the quant (card read in full): no numeric quality figures; card states no formal perplexity, KL-divergence or task evaluation has been completed [L206-207]ba82efc
wtdcode/Qwen3.8-Flash-Next-AWQ-W4A16read in full: no quality claimscompressed-tensors AWQ INT4 (W4A16)No quality claims for the quant (card read in full): card lists only what is quantized [L12-27]0939125
xautonomics/Qwen3.8-Flash-Next-UD-IQ4_XS.LAYR.GGUFread in full: no quality claimsGGUF (per-layer library)No quality claims for the quant (card read in full): support status covers coherence, throughput and token-identity to the monolithic file, not quantization quality [L31-46]9ee54f3
xpashkin/Qwen3.8-Flash-Next-MLX-Serve-iQ-MLX-6.6bpwread in full: no quality claimsMLX (mlx-serve pack)No quality claims for the quant (card read in full): only imatrix correlation checks and allocation error-mass analysis, no output-level quality numbers [L173-177, L193-197]; performance table is speed only [L113-122]fb2a8be
xpashkin/Qwen3.8-Flash-Next-MLX-Serve-iQ-MLX-7.1bpwread in full: no quality claimsMLX (mlx-serve)No quality claims for the quant (card read in full): only throughput/memory; long-context fidelity rationale for bf16 attention is stated without measurements [L73-77, L131-151]f8e2927
Youssofal/Qwen3.8-Flash-Next-MTPLX-Bare-Speedread in full: no quality claimsMLX (MTPLX pack)No quality claims for the quant (card read in full): only speed tables and a sampling-exactness statement for MTP vs plain decoding (not quantization quality) [L62-70, L80-83]4285a4a
Youssofal/Qwen3.8-Flash-Next-MTPLX-Optimized-Qualityread in full: no quality claimsMLX (MTPLX)No quality claims for the quant (card read in full): described as 'highest-fidelity' / 'higher-fidelity' without measured numbers; speed on 256/512 GB Macs not measured [L43, L47, L75-76]56b2d9f
Youssofal/Qwen3.8-Flash-Next-MTPLX-Optimized-Speedread in full: no quality claimsMLX (MTPLX layout)No quality claims for the quant (card read in full): 'Higher quality' vs Bare Speed sibling is asserted without measurements [L45, L111-113]; only speed and speculative-sampling exactness figures [L68-107]34ae535
Zalypamoy91/Huihui-Qwen3.8-Flash-Next-abliterated-GGUFread in full: no quality claimsGGUFmodified weightsNo quality claims for the quant (card read in full): only abliteration description, usage command and warnings [L21-45]8894dad
zerodigest/Qwen3.8-Flash-Next-REAP-320-YMQ-GGUFread in full: no quality claimsGGUFmodified weightsNo quality claims: the perplexity/KLD table lists TBD for both variants [L69-70]0ac7648

How this was built

We listed every repository Hugging Face tags as a quantization of Qwen/Qwen3.8-Flash-Next (453 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 3506 published values appear verbatim on their cited lines (0 failures). For every card that carried measured claims at the first pass (196 of 453), 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 (7). File sizes without a card figure come from Hugging Face's file listing at the same commit.

Not covered: claims made only off-site (other than Unsloth's docs page) or in images other than turboderp's, 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.