Leaderboards, by axis

Every number here comes from the same matrix as the report, with the same rules. Ranked on one axis at a time, because there is no composite score and this project does not publish one — a model that wins code and loses knowledge is not "0.7 of a model".

Three things to read before the first table.

  1. The cap is part of the score. A row measured at 2048 or 4096 tokens per answer is only comparable inside that band. Anything measured below 2048 is marked floor and ranked in its own table at the bottom, never against a ceiling.
  2. n is the item count, not a formality. HumanEval is out of 164, MMLU-Pro out of 120. A row at a smaller n is plumbing and says so.
  3. build is the llama.cpp that produced the row. Decode rates move 10-20% between builds, so a throughput comparison across different builds is a comparison of builds as much as models.

52 measured rows: 33 servable, 19 retired but kept in place, because a leaderboard with the failures deleted is how the same 60 GB gets downloaded twice.


Code — HumanEval, 164 items

#modelHumanEvalncapbuildmeasuredstatus
1lfm25-24b-a2b (thinking)0.9695 (above standard cap)1648192fork b3-1a330952026-11-10✓
2gemma4-31b0.96341642048fork pre-identity2026-10-21✗ retired
3minicpm-v47-35b0.95731644096bolt pre-identity2026-10-18✓
4gemma4-26b-uncen0.95731642048fork b3-1a330952026-10-28✓ driver
5gemma4-26b-a4b0.95731642048fork b3-1a330952026-10-30✗ retired
6ornith-122b0.95121644096fork pre-identity2026-10-20✓
7nemotron-super-120b0.95121644096bolt pre-identity2026-10-18✗ retired
8gemma4-26b-qat-q40.95121642048fork pre-identity2026-10-21✓
9Clover-35B-A3B-Preview (IQ2_M, AR mode)0.9451 (above standard cap)1648192fork b11581-19b458b2026-11-12✓
10mellum21-fork0.9451 (above standard cap)1648192fork b11581-1a330952026-11-20✓
11Nemotron-Bolt-74B-A9B0.94511642048bolt pre-identity2026-10-16✗ retired
12Groot-19B-A3B0.93901644096fork b3-1a330952026-11-11✓
13Axion-1.0-35B (IQ3_XXS)0.9329 (above standard cap)1648192fork b3-1a330952026-11-12✓
14gptoss-abliterated0.92681642048fork pre-identity2026-10-13✓
15ornith-mtp-21g0.92071644096fork pre-identity2026-10-17✓
16flint (i1-Q4_K_M)0.9167604096flint pre-identity2026-10-25✗ retired
17Swift-1.5-FN-Heretic0.9167602048fork b3-1a330952026-11-05✓
18qwen3next-q40.91461642048fork b3-1a330952026-10-31✓
19Ornith-1.5-35B0.90241644096fork b11581-19b458b2026-11-18✓
20Groot-65B-A6B0.9000602048bolt b1-a30fb492026-11-05✗ retired
21lfm25-24b-a2b-htc (thinking)0.89021644096fork b11581-1a330952026-11-16✓
22mellum0.87801644096fork pre-identity2026-10-12✗ retired
23qwen38-35b0.87201644096fork b3-1a330952026-11-01✓
24Laguna-XS-2.10.8667604096fork b3-1a330952026-11-06✓
25Clover-35B-A3B-Preview0.85981644096clover b1-6b9c0cf2026-11-03✗ retired
26lfm25-24b-a2b0.85371644096fork pre-identity2026-10-17✓
27Axion-1.0-35B0.8500602048fork b3-1a330952026-11-05✓
28qwen38fn-abl0.84761642048fork pre-identity2026-10-13✓
29Lite-Beam-M2.7-229B0.8000604096fork b3-1a330952026-11-08✓
30Kimi-Linear-48B-A3B-abl0.7866 (above standard cap)1648192fork b3-1a330952026-11-01✗ retired
31kimi0.78661642048fork pre-identity2026-10-14✗ retired
32hy-memory-30b-a3b0.78051642048fork pre-identity2026-10-17✗ retired
33graft-50b-a3b0.78051642048fork pre-identity2026-10-19✗ retired
34hy-memory-30b-a3b-q80.76831642048fork pre-identity2026-10-17✗ retired
35lfm25-24b-a2b-htc0.68901644096fork b11581-1a330952026-11-16✓
36ornith35b (uncensored)0.56711644096fork b3-1a330952026-10-31✓ boot

The driver is gemma4-26b-uncen — 15.6 GiB, 56.9 t/s, and still the best coder on the box. The three largest models here (122B, 120B) sit below it on code while running 2-2.5× slower: size buys knowledge fidelity on this platform, not code.

Knowledge — MMLU-Pro, 120 items, chat 0-shot

#modelMMLU-Proncapbuildmeasuredstatus
1qwen38fn-abl0.8750402048fork pre-identity2026-10-13✓
2gemma4-31b0.86671202048fork pre-identity2026-10-21✗ retired
3gemma4-26b-a4b0.86671202048fork b3-1a330952026-10-30✗ retired
4gemma4-26b-uncen0.85001202048fork b3-1a330952026-10-28✓ driver
5Clover-35B-A3B-Preview (IQ2_M, AR mode)0.8417 (above standard cap)1208192fork b11581-19b458b2026-11-12✓
6gemma4-26b-qat-q40.84171202048fork pre-identity2026-10-21✓
7mellum21-fork0.8083 (above standard cap)1208192fork b11581-1a330952026-11-20✓
8Swift-1.5-FN-Heretic0.8000402048fork b3-1a330952026-11-05✓
9ornith-122b0.79171204096fork pre-identity2026-10-20✓
10minicpm-v47-35b0.77501204096bolt pre-identity2026-10-18✓
11ornith-mtp-21g0.77501202048fork pre-identity2026-10-17✓
12Axion-1.0-35B0.7750402048fork b3-1a330952026-11-05✓
13nemotron-super-120b0.76671204096bolt pre-identity2026-10-18✗ retired
14Ling-3.0-flash0.76671202048fork pre-identity2026-10-22✗ retired
15Axion-1.0-35B (IQ3_XXS)0.75831204096fork b3-1a330952026-11-12✓
16gptoss-abliterated0.75001202048fork pre-identity2026-10-13✓
17Groot-65B-A6B0.7500402048bolt b1-a30fb492026-11-05✗ retired
18hy-memory-30b-a3b-q80.75001202048fork pre-identity2026-10-17✗ retired
19qwen3next-q40.74171202048fork b3-1a330952026-10-31✓
20hy-memory-30b-a3b0.72501202048fork pre-identity2026-10-17✗ retired
21Groot-19B-A3B0.71671204096fork b3-1a330952026-11-11✓
22Laguna-XS-2.10.7000 (above standard cap)408192fork b3-1a330952026-11-06✓
23mellum0.70001204096fork pre-identity2026-10-12✗ retired
24lfm25-24b-a2b (thinking)0.69171204096fork b3-1a330952026-11-10✓
25lfm25-24b-a2b0.69171204096fork pre-identity2026-10-17✓
26Lite-Beam-M2.7-229B0.6750404096fork b3-1a330952026-11-08✓
27graft-50b-a3b0.65831202048fork pre-identity2026-10-19✗ retired
28qwen38-35b0.64171204096fork b3-1a330952026-11-01✓
29ornith35b (uncensored)0.63331204096fork b3-1a330952026-10-31✓ boot
30lfm25-24b-a2b-htc (thinking)0.61671204096fork b11581-1a330952026-11-16✓
31Kimi-Linear-48B-A3B-abl0.5833 (above standard cap)1208192fork b3-1a330952026-11-01✗ retired
32kimi0.55001202048fork pre-identity2026-10-14✗ retired
33Ornith-1.5-35B0.42501204096fork b11581-19b458b2026-11-18✓
34Clover-35B-A3B-Preview0.40001204096clover b1-6b9c0cf2026-11-03✗ retired
35Nemotron-Bolt-74B-A9B0.36671202048bolt pre-identity2026-10-16✗ retired
36lfm25-24b-a2b-htc0.26671204096fork b11581-1a330952026-11-16✓

The 26B MoE also leads knowledge. The two 120B-class models recover some ground here and none of it justifies 60-70 GiB at 23-27 t/s on a box whose best model fits in a sixth of that.

Reasoning traps — riddles and arithmetic, 14 items

#modelriddlesncapbuildmeasuredstatus
1lfm25-24b-a2b (thinking)14/14——fork b3-1a330952026-11-10✓
2gemma4-31b14/14——fork pre-identity2026-10-21✗ retired
3gemma4-26b-uncen14/14——fork b3-1a330952026-10-28✓ driver
4gemma4-26b-a4b14/14——fork b3-1a330952026-10-30✗ retired
5gemma4-26b-qat-q414/14——fork pre-identity2026-10-21✓
6Clover-35B-A3B-Preview (IQ2_M, AR mode)14/14——fork b11581-19b458b2026-11-12✓
7Axion-1.0-35B (IQ3_XXS)14/14——fork b3-1a330952026-11-12✓
8qwen3next-q614/14——fork pre-identity2026-10-10✓
9Hy4-flash (IQ2_S/IQ2_XXS)14/14——hy4flash-pr b1-f8d4cc42026-11-19✗ retired
10qwen3next-q414/14——fork b3-1a330952026-10-31✓
11Ornith-1.5-35B14/14——fork b11581-19b458b2026-11-18✓
12gemma-4-12B-QAT14/14——fork b3-1a330952026-11-10✓
13lfm25-24b-a2b-htc (thinking)14/14——fork b11581-1a330952026-11-16✓
14qwen38-35b14/14——fork b3-1a330952026-11-01✓
15gemma-4-E4B-QAT14/14——fork b3-1a330952026-11-10✓
16Clover-35B-A3B-Preview14/14——clover b1-6b9c0cf2026-11-03✗ retired
17qwen3-14b14/14——fork b3-1a330952026-11-10✓
18qwen3-30b-a3b14/14——fork b3-1a330952026-11-10✓
19graft-50b-a3b14/14——fork pre-identity2026-10-19✗ retired
20gpt-oss-20b-mxfp414/14——fork b3-1a330952026-11-10✓
21gemma-4-26B-A4B-QAT + MTP drafter14/14——fork b3-1a330952026-11-09✓
22Nemotron-Bolt-74B-A9B13/14——bolt pre-identity2026-10-16✗ retired
23Groot-19B-A3B13/14——fork b3-1a330952026-11-11✓
24Kimi-Linear-48B-A3B-abl13/14——fork b3-1a330952026-11-01✗ retired
25qwen3-8b13/14——fork b3-1a330952026-11-10✓
26kimi13/14——fork pre-identity2026-10-14✗ retired
27ornith-122b12/14——fork pre-identity2026-10-20✓
28mellum21-fork12/14——fork b11581-1a330952026-11-20✓
29gptoss12/14——fork pre-identity2026-10-12✓
30qwen3512/14——fork pre-identity2026-10-11✗ retired
31qwen36-35b12/14——fork pre-identity2026-10-12✗ retired
32mistral-small-3.1-24b12/14——fork b3-1a330952026-11-10✓
33hy-memory-30b-a3b12/14——fork pre-identity2026-10-17✗ retired
34lfm25-24b-a2b-htc12/14——fork b11581-1a330952026-11-16✓
35ornith35b (uncensored)12/14——fork b3-1a330952026-10-31✓ boot
36nemotron-super-120b11/14——bolt pre-identity2026-10-18✗ retired
37ornith-mtp-21g11/14——fork pre-identity2026-10-17✓
38Swift-1.5-FN-Heretic11/14——fork b3-1a330952026-11-05✓
39mellum11/14——fork pre-identity2026-10-12✗ retired
40qwen38fn-abl11/14——fork pre-identity2026-10-13✓
41hy-memory-30b-a3b-q811/14——fork pre-identity2026-10-17✗ retired
42Ling-3.0-flash11/14——fork pre-identity2026-10-22✗ retired
43minicpm-v47-35b10/14——bolt pre-identity2026-10-18✓
44gptoss-abliterated10/14——fork pre-identity2026-10-13✓
45flint (i1-Q4_K_M)10/14——flint pre-identity2026-10-25✗ retired
46Laguna-XS-2.110/14——fork b3-1a330952026-11-06✓
47lfm25-24b-a2b10/14——fork pre-identity2026-10-17✓
48Axion-1.0-35B10/14——fork b3-1a330952026-11-05✓
49Lite-Beam-M2.7-229B10/14——fork b3-1a330952026-11-08✓
50Groot-65B-A6B9/14——bolt b1-a30fb492026-11-05✗ retired
51nemotron-bolt (broken load)8/14——fork pre-identity2026-10-21✗ retired

Fourteen hand-written riddles and three arithmetic traps. Not a benchmark in the psychometric sense: it catches a model that pattern-matches an answer without following the question, which is exactly the failure the code and knowledge axes do not show. 14/14 is the ceiling; several strong coders score 10-11 because they answer confidently and wrongly.

Speed — decode tokens/second

#modeldecodencapbuildmeasuredstatus
1mellum21-fork114.0 t/s——fork b11581-1a330952026-11-20✓
2lfm25-24b-a2b (thinking)111.2 t/s——fork b3-1a330952026-11-10✓
3lfm25-24b-a2b-htc (thinking)109.6 t/s——fork b11581-1a330952026-11-16✓
4lfm25-24b-a2b106.6 t/s——fork pre-identity2026-10-17✓
5lfm25-24b-a2b-htc106.0 t/s——fork b11581-1a330952026-11-16✓
6Axion-1.0-35B (IQ3_XXS)93.8 t/s——fork b3-1a330952026-11-12✓
7mellum91.8 t/s——fork pre-identity2026-10-12✗ retired
8Laguna-XS-2.187.7 t/s——fork b3-1a330952026-11-06✓
9Ornith-1.5-35B80.5 t/s——fork b11581-19b458b2026-11-18✓
10qwen3-30b-a3b79.7 t/s——fork b3-1a330952026-11-10✓
11Axion-1.0-35B76.2 t/s——fork b3-1a330952026-11-05✓
12qwen38-35b74.8 t/s——fork b3-1a330952026-11-01✓
13Kimi-Linear-48B-A3B-abl74.8 t/s——fork b3-1a330952026-11-01✗ retired
14hy-memory-30b-a3b74.7 t/s——fork pre-identity2026-10-17✗ retired
15gpt-oss-20b-mxfp474.4 t/s——fork b3-1a330952026-11-10✓
16Nemotron-Bolt-74B-A9B72.8 t/s——bolt pre-identity2026-10-16✗ retired
17flint (i1-Q4_K_M)71.1 t/s——flint pre-identity2026-10-25✗ retired
18ornith35b (uncensored)69.9 t/s——fork b3-1a330952026-10-31✓ boot
19Clover-35B-A3B-Preview (IQ2_M, AR mode)67.5 t/s——fork b11581-19b458b2026-11-12✓
20kimi67.4 t/s——fork pre-identity2026-10-14✗ retired
21ornith-mtp-21g65.2 t/s——fork pre-identity2026-10-17✓
22Groot-19B-A3B64.8 t/s——fork b3-1a330952026-11-11✓
23minicpm-v47-35b64.6 t/s——bolt pre-identity2026-10-18✓
24gemma4-26b-qat-q459.5 t/s——fork pre-identity2026-10-21✓
25qwen3next-q458.1 t/s——fork b3-1a330952026-10-31✓
26gemma-4-E4B-QAT57.8 t/s——fork b3-1a330952026-11-10✓
27qwen36-35b57.1 t/s——fork pre-identity2026-10-12✗ retired
28gemma4-26b-uncen56.9 t/s——fork b3-1a330952026-10-28✓ driver
29gemma-4-26B-A4B-QAT + MTP drafter54.9 t/s——fork b3-1a330952026-11-09✓
30graft-50b-a3b54.3 t/s——fork pre-identity2026-10-19✗ retired
31Clover-35B-A3B-Preview53.7 t/s——clover b1-6b9c0cf2026-11-03✗ retired
32qwen38fn-abl52.9 t/s——fork pre-identity2026-10-13✓
33hy-memory-30b-a3b-q852.3 t/s——fork pre-identity2026-10-17✗ retired
34gptoss-abliterated50.5 t/s——fork pre-identity2026-10-13✓
35gptoss50.1 t/s——fork pre-identity2026-10-12✓
36gemma4-26b-a4b49.4 t/s——fork b3-1a330952026-10-30✗ retired
37qwen3next-q645.8 t/s——fork pre-identity2026-10-10✓
38Ling-3.0-flash43.5 t/s——fork pre-identity2026-10-22✗ retired
39qwen3-8b39.0 t/s——fork b3-1a330952026-11-10✓
40Lite-Beam-M2.7-229B34.1 t/s——fork b3-1a330952026-11-08✓
41gemma-4-12B-QAT27.8 t/s——fork b3-1a330952026-11-10✓
42Groot-65B-A6B27.6 t/s——bolt b1-a30fb492026-11-05✗ retired
43nemotron-super-120b27.1 t/s——bolt pre-identity2026-10-18✗ retired
44Swift-1.5-FN-Heretic25.6 t/s——fork b3-1a330952026-11-05✓
45qwen3-14b23.5 t/s——fork b3-1a330952026-11-10✓
46ornith-122b23.3 t/s——fork pre-identity2026-10-20✓
47qwen3522.9 t/s——fork pre-identity2026-10-11✗ retired
48nemotron-bolt (broken load)22.7 t/s——fork pre-identity2026-10-21✗ retired
49qwen38fn32-prune220.0 t/s——qwen4exp-var b1-18443252026-11-03✗ retired
50Hy4-flash (IQ2_S/IQ2_XXS)18.9 t/s——hy4flash-pr b1-f8d4cc42026-11-19✗ retired
51gemma4-31b17.0 t/s——fork pre-identity2026-10-21✗ retired
52mistral-small-3.1-24b14.9 t/s——fork b3-1a330952026-11-10✓

Decode is bandwidth-bound on this platform (~256 GB/s shared with the CPU), so it tracks active parameters, not total parameters. The top of this table is two ~12-24B MoEs; the bottom is a 120B.

Prefill — prompt tokens/second

#modelprefillncapbuildmeasuredstatus
1lfm25-24b-a2b (thinking)3360 t/s——fork b3-1a330952026-11-10✓
2lfm25-24b-a2b-htc (thinking)3331 t/s——fork b11581-1a330952026-11-16✓
3lfm25-24b-a2b-htc3225 t/s——fork b11581-1a330952026-11-16✓
4lfm25-24b-a2b2930 t/s——fork pre-identity2026-10-17✓
5mellum21-fork2723 t/s——fork b11581-1a330952026-11-20✓
6mellum2442 t/s——fork pre-identity2026-10-12✗ retired
7gemma-4-E4B-QAT2243 t/s——fork b3-1a330952026-11-10✓
8Groot-19B-A3B2176 t/s——fork b3-1a330952026-11-11✓
9gpt-oss-20b-mxfp42169 t/s——fork b3-1a330952026-11-10✓
10qwen3-30b-a3b1910 t/s——fork b3-1a330952026-11-10✓
11gemma4-26b-qat-q41900 t/s——fork pre-identity2026-10-21✓
12gemma-4-26B-A4B-QAT + MTP drafter1886 t/s——fork b3-1a330952026-11-09✓
13Clover-35B-A3B-Preview (IQ2_M, AR mode)1849 t/s——fork b11581-19b458b2026-11-12✓
14Kimi-Linear-48B-A3B-abl1774 t/s——fork b3-1a330952026-11-01✗ retired
15ornith35b (uncensored)1762 t/s——fork b3-1a330952026-10-31✓ boot
16gemma4-26b-uncen1759 t/s——fork b3-1a330952026-10-28✓ driver
17ornith-mtp-21g1610 t/s——fork pre-identity2026-10-17✓
18minicpm-v47-35b1600 t/s——bolt pre-identity2026-10-18✓
19kimi1597 t/s——fork pre-identity2026-10-14✗ retired
20qwen38-35b1550 t/s——fork b3-1a330952026-11-01✓
21qwen36-35b1548 t/s——fork pre-identity2026-10-12✗ retired
22Axion-1.0-35B (IQ3_XXS)1543 t/s——fork b3-1a330952026-11-12✓
23hy-memory-30b-a3b1515 t/s——fork pre-identity2026-10-17✗ retired
24Ornith-1.5-35B1507 t/s——fork b11581-19b458b2026-11-18✓
25Clover-35B-A3B-Preview1461 t/s——clover b1-6b9c0cf2026-11-03✗ retired
26hy-memory-30b-a3b-q81459 t/s——fork pre-identity2026-10-17✗ retired
27Axion-1.0-35B1422 t/s——fork b3-1a330952026-11-05✓
28graft-50b-a3b1368 t/s——fork pre-identity2026-10-19✗ retired
29Laguna-XS-2.11321 t/s——fork b3-1a330952026-11-06✓
30qwen3-8b1300 t/s——fork b3-1a330952026-11-10✓
31qwen3next-q41219 t/s——fork b3-1a330952026-10-31✓
32qwen3next-q61189 t/s——fork pre-identity2026-10-10✓
33gptoss-abliterated1052 t/s——fork pre-identity2026-10-13✓
34gptoss1045 t/s——fork pre-identity2026-10-12✓
35gemma-4-12B-QAT911 t/s——fork b3-1a330952026-11-10✓
36qwen3-14b743 t/s——fork b3-1a330952026-11-10✓
37qwen35622 t/s——fork pre-identity2026-10-11✗ retired
38ornith-122b605 t/s——fork pre-identity2026-10-20✓
39Groot-65B-A6B597 t/s——bolt b1-a30fb492026-11-05✗ retired
40nemotron-bolt (broken load)586 t/s——fork pre-identity2026-10-21✗ retired
41Nemotron-Bolt-74B-A9B577 t/s——bolt pre-identity2026-10-16✗ retired
42Swift-1.5-FN-Heretic512 t/s——fork b3-1a330952026-11-05✓
43mistral-small-3.1-24b460 t/s——fork b3-1a330952026-11-10✓
44qwen38fn-abl448 t/s——fork pre-identity2026-10-13✓
45nemotron-super-120b411 t/s——bolt pre-identity2026-10-18✗ retired
46qwen38fn32-prune2397 t/s——qwen4exp-var b1-18443252026-11-03✗ retired
47gemma4-31b350 t/s——fork pre-identity2026-10-21✗ retired
48Hy4-flash (IQ2_S/IQ2_XXS)276 t/s——hy4flash-pr b1-f8d4cc42026-11-19✗ retired
49Lite-Beam-M2.7-229B261 t/s——fork b3-1a330952026-11-08✓

Prefill is shader-bound and moves far more between builds than decode does. Use it to size long-prompt work, not to rank models generally.


Floors — HumanEval measured below the 2048-token standard

#modelHumanEvalncapbuildmeasuredstatus
1Hy4-flash (IQ2_S/IQ2_XXS)0.9207 floor164512hy4flash-pr b1-f8d4cc42026-11-19✗ retired
2gemma-4-12B-QAT0.8963 floor164512fork b3-1a330952026-11-10✓
3gemma-4-E4B-QAT0.8659 floor164512fork b3-1a330952026-11-10✓
4qwen3-14b0.8537 floor164512fork b3-1a330952026-11-10✓
5qwen3-30b-a3b0.8537 floor164512fork b3-1a330952026-11-10✓
6mistral-small-3.1-24b0.8232 floor164512fork b3-1a330952026-11-10✓
7qwen3-8b0.7866 floor164512fork b3-1a330952026-11-10✓
8gpt-oss-20b-mxfp40.5854 floor164512fork b3-1a330952026-11-10✓
9nemotron-bolt (broken load)0.1786 floor56512fork pre-identity2026-10-21✗ retired

These rows exist because a floor is still information: it is the lowest a model is known to score. They are not ranked against the tables above, and a plan that depends on one of them should re-measure first.


What is deliberately missing

Source, data and build scripts: github.com/strix-halo-128-benchmarks/strix-halo-128-benchmarks.github.io. Full report and caveats: index.html · deleted models and restore recipes: retired-models.html.