llmrun
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: epoch21 open models ranked+43 proprietaryData through Sep 2026

Open modelsAll models

Open models ranked on ProofBench

# shows rank among open models / rank overall (including proprietary).

#ModelScore
1 / 5Kimi K3 · 2779.9B
87.0%
2 / 13DeepSeek V4 Flash 0731 · 304.2B
56.0%
3 / 18DeepSeek V4 Pro 0813 · 1650.5B
50.0%
4 / 20GLM 5.3 · 753.3B
49.0%
5 / 27GLM 5.2 · 753.3B
35.0%
6 / 32GLM 5.1 · 753.9B
22.2%
7 / 33MiMo V2.5 Pro · 1023.2B
22.0%
8 / 34GLM 5.3 Flash · 321.3B
21.0%
9 / 39MiniMax M3 · 427.0B
18.0%
10 / 41DeepSeek V4 Pro · 1598.8B
16.0%
11 / 42Kimi K2.6 · 1026.9B
16.0%
12 / 43MiMo V2.5 · 310.8B
16.0%
13 / 44Qwen3.8 27B · 27.8B
16.0%
14 / 54DeepSeek V3.2 · 685.4B
8.0%
15 / 55GLM 4.7 · 358.3B
6.0%
16 / 56Inkling Small · 266.0B
6.0%
17 / 59MiniMax M2.5 · 228.7B
4.0%
18 / 60MiniMax M2.7 · 228.7B
3.0%
19 / 62Inkling · 952.4B
0.0%
20 / 63Laguna M.1 · 225.8B
0.0%
21 / 64Laguna XS.2 · 33.4B
0.0%

Score vs model size

Which models give the most quality for their size — the ones worth running locally.

100B1Tmodel size (log scale) →87.0%0.0%DeepSeek V4 Pro 0813 · 1.7T · 50.0%GLM 5.3 · 753B · 49.0%GLM 5.2 · 753B · 35.0%GLM 5.1 · 754B · 22.2%MiMo V2.5 Pro · 1T · 22.0%GLM 5.3 Flash · 321B · 21.0%MiniMax M3 · 427B · 18.0%DeepSeek V4 Pro · 1.6T · 16.0%Kimi K2.6 · 1T · 16.0%MiMo V2.5 · 311B · 16.0%DeepSeek V3.2 · 685B · 8.0%Inkling Small · 266B · 6.0%GLM 4.7 · 358B · 6.0%MiniMax M2.5 · 229B · 4.0%MiniMax M2.7 · 229B · 3.0%Laguna M.1 · 226B · 0.0%Laguna XS.2 · 33B · 0.0%Inkling · 952B · 0.0%Qwen3.8 27B · 28B · 16.0%Qwen3.8 27BDeepSeek V4 Flash 0731 · 304B · 56.0%DeepSeek V4 Flash 0731Kimi K3 · 2.8T · 87.0%Kimi K3
Each dot is a model. Up = higher score, left = smaller (easier to run locally). The dashed line marks the efficiency frontier — the best score you can get at each size or smaller.
  • 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).
  • 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 #5. 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 (about 15 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.