Reasoning
CritPt Leaderboard
CritPt (Critical Points) tests graduate-level physics problem-solving on research-style questions spanning mechanics, electromagnetism, quantum mechanics, and more. It was created by physicists to probe whether a model can reason through multi-step physics rather than pattern-match textbook answers.
Source: epoch66 open models ranked+153 proprietaryData through Oct 2026
Open models ranked on CritPt
# shows rank among open models / rank overall (including proprietary).
| # | Model | Score | VRAM @ Q4 |
|---|---|---|---|
| 1 / 36 | MiMo V2.6 Pro RL · 1020B | 26.6% | ~613 GB |
| 2 / 48 | Kimi K3 · 2779.9B | 23.4% | ~1674 GB |
| 3 / 54 | GLM 5.2 · 753.3B | 20.9% | ~456 GB |
| 4 / 62 | GLM 5.3 · 753.3B | 19.1% | ~456 GB |
| 5 / 67 | DeepSeek V4 Pro 0813 · 1650.5B | 18.0% | ~991 GB |
| 6 / 80 | DeepSeek V4 Flash 0731 · 304.2B | 16.6% | ~183 GB |
| 7 / 84 | GLM 5.3 Flash · 321.3B | 15.4% | ~195 GB |
| 8 / 90 | DeepSeek V4.1 Flash · 763.2B | 14.3% | ~458 GB |
| 9 / 96 | DeepSeek V4 Pro · 1598.8B | 12.9% | ~960 GB |
| 10 / 100 | MiMo V2.6 Flash RL · 309B | 12.0% | ~186 GB |
| 11 / 107 | Kimi K2.7 Code · 1026.9B | 10.0% | ~620 GB |
| 12 / 115 | Inkling Small · 266.0B | 8.3% | ~160 GB |
| 13 / 118 | Kimi K2.6 · 1026.9B | 8.0% | ~620 GB |
| 14 / 119 | DeepSeek V4 Flash · 290.9B | 7.1% | ~175 GB |
| 15 / 123 | Inkling · 952.4B | 5.4% | ~572 GB |
| 16 / 124 | Qwen3.8 27B · 27.8B | 5.4% | ~17 GB |
| 17 / 129 | GLM 5.1 · 753.9B | 4.6% | ~457 GB |
| 18 / 132 | MiMo V2.5 Pro · 1023.2B | 4.0% | ~614 GB |
| 19 / 133 | MiMo V2.5 · 310.8B | 3.7% | ~187 GB |
| 20 / 134 | MiniMax M3 · 427.0B | 3.7% | ~257 GB |
| 21 / 135 | Ring 2.6 1T · 1025.4B | 3.7% | ~2057 GB |
| 22 / 137 | Kimi K2.5 · 1026.9B | 3.1% | ~620 GB |
| 23 / 138 | NVIDIA Nemotron 3 Super 120B A12B BF16 · 123.6B | 3.1% | ~75 GB |
| 24 / 139 | NVIDIA Nemotron 3 Ultra 550B A55B BF16 · 560.5B | 3.1% | ~370 GB |
| 25 / 140 | DeepSeek V3.2 Exp · 685.4B | 2.9% | ~415 GB |
| 26 / 144 | Muse Glimmer 30B · 29.8B | 2.6% | ~18 GB |
| 27 / 145 | Step 3.7 Flash · 201.4B | 2.3% | ~133 GB |
| 28 / 148 | DeepSeek V3.1 Terminus · 684.5B | 1.7% | ~415 GB |
| 29 / 149 | GLM 4.7 · 358.3B | 1.7% | ~216 GB |
| 30 / 150 | Gemma 4 31B IT · 31.3B | 1.4% | ~20 GB |
| 31 / 152 | GPT OSS 20B · 20.9B | 1.4% | ~13 GB |
| 32 / 157 | GLM 4.6 · 356.8B | 1.1% | ~215 GB |
| 33 / 159 | GPT OSS 120B · 116.8B | 1.1% | ~71 GB |
| 34 / 160 | DeepSeek R1 · 684.5B | 1.1% | ~415 GB |
| 35 / 162 | Qwen3.5 122B A10B · 125.1B | 0.9% | ~75 GB |
| 36 / 163 | Qwen3.6 27B · 27.8B | 0.9% | ~17 GB |
| 37 / 164 | Trinity Large Thinking · 398.6B | 0.9% | ~798 GB |
| 38 / 167 | MiniMax M2.7 · 228.7B | 0.6% | ~138 GB |
| 39 / 168 | Qwen3.5 35B A3B · 36.0B | 0.6% | ~22 GB |
| 40 / 171 | Command A Plus 05 2026 BF16 · 218.8B | 0.3% | ~132 GB |
| 41 / 174 | Magistral Small 2509 · 24.0B | 0.3% | ~15 GB |
| 42 / 176 | Qwen3 30B A3B Thinking 2507 · 30.5B | 0.3% | ~19 GB |
| 43 / 177 | Qwen3 32B · 32.8B | 0.3% | ~20 GB |
| 44 / 178 | Qwen3.5 9B · 9.7B | 0.3% | ~6.4 GB |
| 45 / 179 | Qwen3.6 35B A3B · 36.0B | 0.3% | ~22 GB |
| 46 / 182 | Cogito 671B V2.1 · 671B | 0.0% | ~406 GB |
| 47 / 183 | DeepSeek v3 · 684.5B | 0.0% | ~415 GB |
| 48 / 184 | DeepSeek v3 0324 · 684.5B | 0.0% | ~415 GB |
| 49 / 185 | Devstral Small 2 24B Instruct 2512 · 24.0B | 0.0% | ~15 GB |
| 50 / 187 | Gemma 3 12B IT · 12.2B | 0.0% | ~8.0 GB |
| 51 / 188 | Gemma 3 27B IT · 27.4B | 0.0% | ~18 GB |
| 52 / 189 | Gemma 4 26B A4B IT · 25.8B | 0.0% | ~16 GB |
| 53 / 197 | Granite 4.1 30B · 28.9B | 0.0% | ~18 GB |
| 54 / 200 | Llama 3.1 8B Instruct · 8.0B | 0.0% | ~5.3 GB |
| 55 / 201 | Llama 3.3 70B Instruct · 70.6B | 0.0% | ~47 GB |
| 56 / 202 | Llama 4 Maverick 17B 128E Instruct · 401.6B | 0.0% | ~265 GB |
| 57 / 203 | Llama 4 Scout 17B 16E Instruct · 108.6B | 0.0% | ~72 GB |
| 58 / 205 | MiMo v2 Flash · 309.8B | 0.0% | ~186 GB |
| 59 / 206 | Mistral Large 3 675B Instruct 2512 · 675B | 0.0% | ~446 GB |
| 60 / 209 | Mistral Small 3.1 24B Instruct 2503 · 24.0B | 0.0% | ~15 GB |
| 61 / 210 | Mistral Small 3.2 24B Instruct 2506 · 24.0B | 0.0% | ~15 GB |
| 62 / 214 | Phi 4 Mini Instruct · 3.8B | 0.0% | ~2.9 GB |
| 63 / 215 | Qwen3 14B · 14.8B | 0.0% | ~9.5 GB |
| 64 / 216 | Qwen3 235B A22B Thinking 2507 · 235.1B | 0.0% | ~142 GB |
| 65 / 217 | Qwen3 8B · 8.2B | 0.0% | ~5.5 GB |
| 66 / 218 | Qwen3 Coder Next · 79.7B | 0.0% | ~48 GB |
Score vs model size
Which models give the most quality for their size — the ones worth running locally.
- Phi 4 Mini Instruct, 4B, score 0.0% — on the efficiency frontier (best score at its size or smaller).
- Qwen3.5 9B, 10B, score 0.3% — on the efficiency frontier (best score at its size or smaller).
- GPT OSS 20B, 21B, score 1.4% — on the efficiency frontier (best score at its size or smaller).
- Qwen3.8 27B, 28B, score 5.4% — on the efficiency frontier (best score at its size or smaller).
- Inkling Small, 266B, score 8.3% — on the efficiency frontier (best score at its size or smaller).
- DeepSeek V4 Flash 0731, 304B, score 16.6% — on the efficiency frontier (best score at its size or smaller).
- GLM 5.2, 753B, score 20.9% — on the efficiency frontier (best score at its size or smaller).
- MiMo V2.6 Pro RL, 1T, score 26.6% — on the efficiency frontier (best score at its size or smaller).
CritPt: frequently asked questions
- What is the best open LLM on CritPt?
- MiMo V2.6 Pro RL is the top open model on CritPt, scoring 26.6%. Among all models tested — including proprietary ones — it ranks #36. The top model overall is GPT 5.6 Sol Max (OpenAI) at 32.3%.
- What's the best CritPt 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 5.4% on CritPt.
- What's the best CritPt model you can run on a 12 GB GPU?
- Qwen3.5 9B is the highest-scoring open model that fits in 12 GB at 4-bit quantization (~6.4 GB), scoring 0.3% on CritPt.
- Can open models match proprietary models on CritPt?
- Not quite on CritPt: the strongest proprietary model (GPT 5.6 Sol Max) scores 32.3%, ahead of the best open model (MiMo V2.6 Pro RL) at 26.6% — 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.