Coding
SciCode Leaderboard
SciCode asks a model to write real scientific-computing code — implementing multi-step numerical methods drawn from physics, chemistry, biology, and math research — and checks the output against reference solutions. It was built by a team of scientists and ML researchers to test coding ability on genuine research problems, not typical software tasks.
Source: epoch66 open models ranked+141 proprietaryData through Oct 2026
Open models ranked on SciCode
# shows rank among open models / rank overall (including proprietary).
| # | Model | Score | VRAM @ Q4 |
|---|---|---|---|
| 1 / 8 | MiMo V2.6 Pro RL · 1020B | 60.9% | ~613 GB |
| 2 / 13 | Kimi K3 · 2779.9B | 59.5% | ~1674 GB |
| 3 / 15 | GLM 5.3 · 753.3B | 59.0% | ~456 GB |
| 4 / 75 | Kimi K2.6 · 1026.9B | 53.5% | ~620 GB |
| 5 / 87 | DeepSeek V4.1 Flash · 763.2B | 51.8% | ~458 GB |
| 6 / 91 | GLM 5.3 Flash · 321.3B | 51.6% | ~195 GB |
| 7 / 95 | MiMo V2.6 Flash RL · 309B | 51.3% | ~186 GB |
| 8 / 96 | GLM 5.2 · 753.3B | 51.2% | ~456 GB |
| 9 / 97 | DeepSeek V4 Pro 0813 · 1650.5B | 51.0% | ~991 GB |
| 10 / 105 | MiMo V2.5 Pro · 1023.2B | 50.2% | ~614 GB |
| 11 / 107 | MiniMax M2.7 · 228.7B | 50.1% | ~138 GB |
| 12 / 108 | DeepSeek V4 Pro · 1598.8B | 50.0% | ~960 GB |
| 13 / 109 | DeepSeek V4 Flash 0731 · 304.2B | 49.9% | ~183 GB |
| 14 / 111 | Inkling Small · 266.0B | 49.6% | ~160 GB |
| 15 / 117 | Kimi K2.5 · 1026.9B | 49.0% | ~620 GB |
| 16 / 121 | Kimi K2.7 Code · 1026.9B | 47.8% | ~620 GB |
| 17 / 126 | MiniMax M3 · 427.0B | 47.1% | ~257 GB |
| 18 / 127 | Inkling · 952.4B | 47.0% | ~572 GB |
| 19 / 131 | Qwen3.8 27B · 27.8B | 46.6% | ~17 GB |
| 20 / 136 | Gemma 4 31B IT · 31.3B | 45.5% | ~20 GB |
| 21 / 137 | GLM 4.7 · 358.3B | 45.1% | ~216 GB |
| 22 / 139 | Ring 2.6 1T · 1025.4B | 45.0% | ~2057 GB |
| 23 / 140 | DeepSeek V4 Flash · 290.9B | 44.9% | ~175 GB |
| 24 / 141 | Muse Glimmer 30B · 29.8B | 44.9% | ~18 GB |
| 25 / 142 | GLM 5.1 · 753.9B | 44.8% | ~457 GB |
| 26 / 146 | MiMo V2.5 · 310.8B | 43.1% | ~187 GB |
| 27 / 151 | Qwen3 235B A22B Thinking 2507 · 235.1B | 41.4% | ~142 GB |
| 28 / 153 | Cogito 671B V2.1 · 671B | 41.0% | ~406 GB |
| 29 / 157 | NVIDIA Nemotron 3 Ultra 550B A55B BF16 · 560.5B | 40.3% | ~370 GB |
| 30 / 160 | Gemma 4 26B A4B IT · 25.8B | 40.1% | ~16 GB |
| 31 / 161 | Step 3.7 Flash · 201.4B | 40.1% | ~133 GB |
| 32 / 164 | DeepSeek v3 0324 · 684.5B | 39.0% | ~415 GB |
| 33 / 166 | DeepSeek V3.2 Exp · 685.4B | 38.9% | ~415 GB |
| 34 / 169 | Command A Plus 05 2026 BF16 · 218.8B | 38.5% | ~132 GB |
| 35 / 171 | GLM 4.6 · 356.8B | 38.4% | ~215 GB |
| 36 / 172 | DeepSeek R1 · 684.5B | 38.3% | ~415 GB |
| 37 / 173 | DeepSeek V3.1 Terminus · 684.5B | 38.0% | ~415 GB |
| 38 / 174 | Qwen3.6 27B · 27.8B | 37.3% | ~17 GB |
| 39 / 175 | Mistral Large 3 675B Instruct 2512 · 675B | 36.6% | ~446 GB |
| 40 / 176 | Qwen3.6 35B A3B · 36.0B | 36.6% | ~22 GB |
| 41 / 177 | Trinity Large Thinking · 398.6B | 36.1% | ~798 GB |
| 42 / 178 | GPT OSS 120B · 116.8B | 36.0% | ~71 GB |
| 43 / 179 | NVIDIA Nemotron 3 Super 120B A12B BF16 · 123.6B | 36.0% | ~75 GB |
| 44 / 180 | DeepSeek v3 · 684.5B | 35.8% | ~415 GB |
| 45 / 181 | Qwen3.5 122B A10B · 125.1B | 35.6% | ~75 GB |
| 46 / 182 | Qwen3 32B · 32.8B | 35.4% | ~20 GB |
| 47 / 183 | Magistral Small 2509 · 24.0B | 35.2% | ~15 GB |
| 48 / 184 | GPT OSS 20B · 20.9B | 34.4% | ~13 GB |
| 49 / 185 | Llama 4 Maverick 17B 128E Instruct · 401.6B | 33.1% | ~265 GB |
| 50 / 186 | Qwen3 30B A3B Thinking 2507 · 30.5B | 33.0% | ~19 GB |
| 51 / 188 | Qwen3 Coder Next · 79.7B | 32.3% | ~48 GB |
| 52 / 189 | Qwen3 14B · 14.8B | 31.6% | ~9.5 GB |
| 53 / 190 | Qwen3.5 9B · 9.7B | 29.5% | ~6.4 GB |
| 54 / 191 | Qwen3.5 35B A3B · 36.0B | 29.3% | ~22 GB |
| 55 / 192 | Devstral Small 2 24B Instruct 2512 · 24.0B | 28.8% | ~15 GB |
| 56 / 194 | Mistral Small 3.1 24B Instruct 2503 · 24.0B | 27.8% | ~15 GB |
| 57 / 196 | Mistral Small 3.2 24B Instruct 2506 · 24.0B | 26.4% | ~15 GB |
| 58 / 197 | Llama 3.3 70B Instruct · 70.6B | 26.0% | ~47 GB |
| 59 / 199 | MiMo v2 Flash · 309.8B | 25.9% | ~186 GB |
| 60 / 200 | Granite 4.1 30B · 28.9B | 25.8% | ~18 GB |
| 61 / 202 | Gemma 3 27B IT · 27.4B | 23.3% | ~18 GB |
| 62 / 203 | Qwen3 8B · 8.2B | 22.6% | ~5.5 GB |
| 63 / 204 | Llama 4 Scout 17B 16E Instruct · 108.6B | 17.0% | ~72 GB |
| 64 / 205 | Gemma 3 12B IT · 12.2B | 16.4% | ~8.0 GB |
| 65 / 206 | Llama 3.1 8B Instruct · 8.0B | 13.2% | ~5.3 GB |
| 66 / 207 | Phi 4 Mini Instruct · 3.8B | 10.8% | ~2.9 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 10.8% — on the efficiency frontier (best score at its size or smaller).
- Llama 3.1 8B Instruct, 8B, score 13.2% — on the efficiency frontier (best score at its size or smaller).
- Qwen3 8B, 8B, score 22.6% — on the efficiency frontier (best score at its size or smaller).
- Qwen3.5 9B, 10B, score 29.5% — on the efficiency frontier (best score at its size or smaller).
- Qwen3 14B, 15B, score 31.6% — on the efficiency frontier (best score at its size or smaller).
- GPT OSS 20B, 21B, score 34.4% — on the efficiency frontier (best score at its size or smaller).
- Magistral Small 2509, 24B, score 35.2% — on the efficiency frontier (best score at its size or smaller).
- Gemma 4 26B A4B IT, 26B, score 40.1% — on the efficiency frontier (best score at its size or smaller).
- Qwen3.8 27B, 28B, score 46.6% — on the efficiency frontier (best score at its size or smaller).
- MiniMax M2.7, 229B, score 50.1% — on the efficiency frontier (best score at its size or smaller).
- MiMo V2.6 Flash RL, 309B, score 51.3% — on the efficiency frontier (best score at its size or smaller).
- GLM 5.3 Flash, 321B, score 51.6% — on the efficiency frontier (best score at its size or smaller).
- GLM 5.3, 753B, score 59.0% — on the efficiency frontier (best score at its size or smaller).
- MiMo V2.6 Pro RL, 1T, score 60.9% — on the efficiency frontier (best score at its size or smaller).
SciCode: frequently asked questions
- What is the best open LLM on SciCode?
- MiMo V2.6 Pro RL is the top open model on SciCode, scoring 60.9%. Among all models tested — including proprietary ones — it ranks #7. The top model overall is Claude Opus 5.5 Max (Anthropic) at 66.9%.
- What's the best SciCode 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 46.6% on SciCode.
- What's the best SciCode model you can run on a 12 GB GPU?
- Qwen3 14B is the highest-scoring open model that fits in 12 GB at 4-bit quantization (~9.5 GB), scoring 31.6% on SciCode.
- Can open models match proprietary models on SciCode?
- Not quite on SciCode: the strongest proprietary model (Claude Opus 5.5 Max) scores 66.9%, ahead of the best open model (MiMo V2.6 Pro RL) at 60.9% — 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.