Skip to content

Scientific Code Benchmark

SciCode evaluates language models on generating code for realistic scientific research problems across 16 subfields of physics, math, chemistry, biology, and material science. Problems decompose into 338 subproblems requiring domain knowledge recall, scientific reasoning, and precise code synthesis. Based on real scripts from published research.

23 models

Higher is better.

  1. 1Step 5 PreviewStepFun58.9
  2. 2Qwen3.7 MaxAlibaba53.5
  3. 3Gemini 3.5 FlashGoogle53.1
  4. 4Kimi K2.6Moonshot AI52.2
  5. 5Qwen3.7 PlusAlibaba51.3
  6. 6Beam 1.0 501B-A23BReflection AI49.2
  7. 7Kimi K2.5Moonshot AI48.7
  8. 8Inkling-SmallThinking Machines Lab48.7
  9. 9Grok 4.3xAI47.3
  10. 10Qwen 3.6 Max (preview)Alibaba47.0
  11. 11Nemotron 3 UltraNVIDIA44.6
  12. 12Muse Glimmer 30BMeta43.6
  13. 13Ling 3.0 FlashInclusionAI41.2
  14. 14Hy3 PreviewTencent41.2
  15. 15A.X K2SK Telecom41.0
  16. 16Mercury 2.5Inception38.0
  17. 17K-EXAONE 2.0LG AI Research37.4
  18. 18Granite 4.2 8BIBM36.1
  19. 19LFM2.5-24B-A2BLiquid AI32.7
  20. 20Nemotron 3 Nano Omni 30B A3BNVIDIA32.0
  21. 21Nemotron 3.5 Lightning 30B A3B NVFP4NVIDIA31.4
  22. 22Ling 2.6 FlashInclusionAI27.0
  23. 23MiniCPM5-2BOpenBMB26.3
Higher is better. Bar colour: blue proprietary, orange open weights, green other.

About this benchmark

Tasks
80

Scores are as reported by model developers or evaluators and may use different settings (for example reasoning effort or tools). Compare with care.