Math
ProofBench Leaderboard
ProofBench evaluates a model's ability to write and check rigorous mathematical proofs, rather than just produce a final numeric answer. It probes step-by-step logical correctness, which is a harder and more distinct skill than solving for a number.
Source: epoch26 open models ranked+56 proprietaryData through Oct 2026
Open models ranked on ProofBench
# shows rank among open models / rank overall (including proprietary).
| # | Model | Score | VRAM @ Q4 |
|---|---|---|---|
| 1 / 10 | Kimi K3 · 2779.9B | 87.0% | ~1674 GB |
| 2 / 15 | Hy4 Preview · 780.0B | 75.0% | ~1561 GB |
| 3 / 17 | MiMo V2.6 Pro RL · 1020B | 70.0% | ~613 GB |
| 4 / 20 | MiMo V2.6 Flash RL · 309B | 63.0% | ~186 GB |
| 5 / 25 | DeepSeek V4 Flash 0731 · 304.2B | 56.0% | ~183 GB |
| 6 / 29 | DeepSeek V4.1 Flash · 763.2B | 54.0% | ~458 GB |
| 7 / 32 | DeepSeek V4 Pro 0813 · 1650.5B | 50.0% | ~991 GB |
| 8 / 34 | GLM 5.3 · 753.3B | 49.0% | ~456 GB |
| 9 / 42 | GLM 5.2 · 753.3B | 35.0% | ~456 GB |
| 10 / 48 | GLM 5.1 · 753.9B | 22.2% | ~457 GB |
| 11 / 49 | MiMo V2.5 Pro · 1023.2B | 22.0% | ~614 GB |
| 12 / 50 | GLM 5.3 Flash · 321.3B | 21.0% | ~195 GB |
| 13 / 55 | MiniMax M3 · 427.0B | 18.0% | ~257 GB |
| 14 / 57 | DeepSeek V4 Pro · 1598.8B | 16.0% | ~960 GB |
| 15 / 58 | Kimi K2.6 · 1026.9B | 16.0% | ~620 GB |
| 16 / 59 | MiMo V2.5 · 310.8B | 16.0% | ~187 GB |
| 17 / 60 | Qwen3.8 27B · 27.8B | 16.0% | ~17 GB |
| 18 / 71 | DeepSeek V3.2 · 685.4B | 8.0% | ~415 GB |
| 19 / 72 | GLM 4.7 · 358.3B | 6.0% | ~216 GB |
| 20 / 73 | Inkling Small · 266.0B | 6.0% | ~160 GB |
| 21 / 76 | MiniMax M2.5 · 228.7B | 4.0% | ~138 GB |
| 22 / 78 | MiniMax M2.7 · 228.7B | 3.0% | ~138 GB |
| 23 / 79 | NVIDIA Nemotron 3 Ultra 550B A55B BF16 · 560.5B | 2.0% | ~370 GB |
| 24 / 80 | Inkling · 952.4B | 0.0% | ~572 GB |
| 25 / 81 | Laguna M.1 · 225.8B | 0.0% | ~136 GB |
| 26 / 82 | Laguna XS.2 · 33.4B | 0.0% | ~20 GB |
Score vs model size
Which models give the most quality for their size — the ones worth running locally.
- Qwen3.8 27B, 28B, score 16.0% — on the efficiency frontier (best score at its size or smaller).
- DeepSeek V4 Flash 0731, 304B, score 56.0% — on the efficiency frontier (best score at its size or smaller).
- MiMo V2.6 Flash RL, 309B, score 63.0% — on the efficiency frontier (best score at its size or smaller).
- Hy4 Preview, 780B, score 75.0% — on the efficiency frontier (best score at its size or smaller).
- Kimi K3, 2.8T, score 87.0% — on the efficiency frontier (best score at its size or smaller).
ProofBench: frequently asked questions
- What is the best open LLM on ProofBench?
- Kimi K3 is the top open model on ProofBench, scoring 87.0%. Among all models tested — including proprietary ones — it ranks #10. The top model overall is Claude Fable 5.1 Max (Anthropic) at 100.0%.
- What's the best ProofBench model you can run on a 24 GB GPU?
- Qwen3.8 27B is the highest-scoring open model that fits in 24 GB at 4-bit quantization (~17 GB), scoring 16.0% on ProofBench.
- Can open models match proprietary models on ProofBench?
- Not quite on ProofBench: the strongest proprietary model (Claude Fable 5.1 Max) scores 100.0%, ahead of the best open model (Kimi K3) at 87.0% — but you can run the open one yourself.
Scores aggregated from epoch. llmrun does not run this benchmark — see the source for methodology, or the about benchmarks for what it measures.