Coding · Benchmark profile
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.
Data verified 27 Jul 2026 · Methodology 1.8.0
Benchmark score on Scientific Code Benchmark
Sakana Fugu leads at 60.1%, followed by Sakana Fugu-Ultra (58.7%) and Qwen3.7-Max (53.5%).
10 ranked models · higher is better · labels show rank and score
| Model | Rank | Provider | Score |
|---|---|---|---|
| Sakana Fugu | #1 | Sakana AI | 60.1% |
| Sakana Fugu-Ultra | #2 | Sakana AI | 58.7% |
| Qwen3.7-Max | #3 | Alibaba Cloud | 53.5% |
| Gemini 3.5 Flash | #4 | 53.1% | |
| Kimi K2.6 | #5 | Moonshot AI | 52.2% |
| Qwen3.7-Plus | #6 | Alibaba Cloud | 51.3% |
| Kimi K2.5 | #7 | Moonshot AI | 48.7% |
| Grok 4.3 | #8 | xAI | 47.3% |
| Qwen 3.6 Max (preview) | #9 | Alibaba Cloud | 47% |
| Nemotron 3 Ultra | #10 | NVIDIA | 44.6% |
One best score per model · higher is better
| Rank | Model | Provider | License | Evidence use | Score |
|---|---|---|---|---|---|
| #1 | Sakana Fugu sakana-fugu | Sakana AI | closed | Estimated reference | 60.1% |
| #2 | Sakana Fugu-Ultra sakana-fugu-ultra | Sakana AI | closed | Estimated reference | 58.7% |
| #3 | Qwen3.7-Max qwen3.7-max | Alibaba Cloud | closed | Reference only | 53.5% |
| #4 | Gemini 3.5 Flash gemini-3.5-flash | closed | Estimated reference | 53.1% | |
| #5 | Kimi K2.6 kimi-k2.6 | Moonshot AI | open | Estimated reference | 52.2% |
| #6 | Qwen3.7-Plus qwen3.7-plus | Alibaba Cloud | closed | Reference only | 51.3% |
| #7 | Kimi K2.5 kimi-k2.5 | Moonshot AI | open | Reference only | 48.7% |
| #8 | Grok 4.3 grok-4.3 | xAI | closed | Reference only | 47.3% |
| #9 | Qwen 3.6 Max (preview) qwen3-6-max-preview | Alibaba Cloud | closed | Reference only | 47% |
| #10 | Nemotron 3 Ultra nemotron-3-ultra-550b | NVIDIA | open | Estimated reference | 44.6% |
| #11 | Hy3 Preview hy3-preview | Tencent | open | Estimated reference | 41.2% |
| #12 | Nemotron 3 Nano Omni 30B A3B nemotron-3-nano-omni-30b-a3b | NVIDIA | open | Estimated reference | 32% |
| #13 | Ling 2.6 Flash ling-2-6-flash | InclusionAI | open | Estimated reference | 27% |
The top of this snapshot is led by Sakana Fugu at 60.1%; third place is 6.6 points behind. The top-10 spread is 15.5 points.
About 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. Results stay tied to the exact model variant and evaluation system. Multiple systems for the same model use the best published score on this page; overall Lumina scoring uses the median of ranking-eligible rows.
Open benchmark source ↗FAQ
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.
Sakana Fugu by Sakana AI currently leads with 60.1%.
13 models in the LuminaBench cohort have a qualifying score on this benchmark.
Related
