llmrun Composite Score · Science
The best open LLMs for science, ranked on graduate-level science and knowledge boards such as GPQA. The score shown is the overall llmrun Composite Score; the highlighted column is the scientific knowledge skill score used for the order.
40 benchmarks4 sourcesUpdated 10 Oct 2026Reference scale 2026-Q4How the Composite Score is built
Science score vs VRAM needed
This chart doesn't show proprietary models.
| # (open / overall) | Model | llmrun Composite Score | Coding | Agents | Math | Science | Reasoning | Benchmarks | VRAM |
|---|---|---|---|---|---|---|---|---|---|
| 1 / 67 | Qwen3.8 27B17.9 GB · 12 benchmarks | 112 | – | 118 | 112 | 124 | 12 | 17.9 GB | |
| 2 / 72 | Qwen3.6 27B17.9 GB · 12 benchmarks | 102 | – | 114 | 111 | 106 | 12 | 17.9 GB | |
| 3 / 75 | Qwen3.6 35B A3B22.0 GB · 12 benchmarks | 105 | – | 109 | 110 | 107 | 12 | 22.0 GB | |
| 4 / 80 | Qwen3.5 35B A3B22.0 GB · 8 benchmarks | – | – | – | 108 | 106 | 8 | 22.0 GB | |
| 5 / 86 | Gemma 4 31B IT21.7 GB · 11 benchmarks | 108 | – | – | 105 | 109 | 11 | 21.7 GB | |
| 6 / 88 | Qwen3.5 9B6.6 GB · 9 benchmarks | – | – | – | 104 | 97 | 9 | 6.6 GB | |
| 7 / 94 | Gemma 4 26B A4B IT16.5 GB · 9 benchmarks | 104 | – | – | 102 | 101 | 9 | 16.5 GB | |
| 8 / 101 | Qwen3 30B A3B Thinking 250719.4 GB · 8 benchmarks | – | – | – | 99 | 94 | 8 | 19.4 GB | |
| 9 / 113 | Qwen3 32B22.6 GB · 9 benchmarks | 93 | – | – | 96 | 91 | 9 | 22.6 GB | |
| 10 / 117 | Qwen3 14B11.9 GB · 8 benchmarks | – | – | – | 94 | 88 | 8 | 11.9 GB | |
| 11 / 119 | GPT OSS 20B13.1 GB · 11 benchmarks | 101 | – | – | 92 | 91 | 11 | 13.1 GB | |
| 12 / 126 | Qwen3 8B7.6 GB · 8 benchmarks | – | – | – | 88 | 81 | 8 | 7.6 GB | |
| 13 / 127 | Devstral Small 2 24B Instruct 251218.1 GB · 3 benchmarks | — | – | – | – | 86 | – | 3 | 18.1 GB |
| 14 / 129 | Magistral Small 250918.1 GB · 6 benchmarks | – | – | – | 83 | 85 | 6 | 18.1 GB | |
| 15 / 130 | Mistral Small 3.2 24B Instruct 250618.1 GB · 6 benchmarks | – | – | – | 82 | 83 | 6 | 18.1 GB | |
| 16 / 133 | Granite 4.1 30B21.9 GB · 2 benchmarks | — | – | – | – | 82 | – | 2 | 21.9 GB |
| 17 / 135 | Mistral Small 3.1 24B Instruct 250318.1 GB · 6 benchmarks | – | – | – | 81 | 82 | 6 | 18.1 GB | |
| 18 / 138 | Gemma 3 27B IT21.7 GB · 10 benchmarks | 69 | – | – | 80 | 74 | 10 | 21.7 GB | |
| 19 / 140 | Gemma 3 12B IT10.9 GB · 8 benchmarks | – | – | – | 72 | 66 | 8 | 10.9 GB | |
| 20 / 141 | Phi 4 Mini Instruct4.8 GB · 3 benchmarks | — | – | – | – | 61 | – | 3 | 4.8 GB |
| 21 / 142 | Llama 3.1 8B Instruct7.9 GB · 10 benchmarks | 56 | – | – | 56 | 69 | 10 | 7.9 GB |
The range after each score is a 90% interval: given the boards a model has results on, its true score is very likely inside it. Short ranges mean many agreeing results; long ranges mean few or conflicting ones.
Why some models are missing. A model receives a score only if it has results on at least 4 benchmarks across at least 2 categories, with one being reasoning or coding. Skill scores require at least 2 benchmarks for that skill; otherwise they display as –.
VRAM is llmrun's estimate at Q4_K_M (or the smallest quantization we track) for the weights plus room for a 16K-token context. Longer contexts need more memory. Proprietary models can't run locally, so the VRAM filter hides them.
How the llmrun Composite Score is built · llmrun doesn't run these benchmarks; scores come from public sources.