SWE-rebench
Code · 2025-05-14
SWE-rebench is a continuously updated, decontaminated software-engineering benchmark exposed through a dynamic public leaderboard. The parent benchmark stores the resolved-rate score over defined evaluation windows; pass@k variants are stored as child benchmark metrics.
Top models (higher is better)
| Model | Score |
|---|---|
| MiniMax M3 | 47.2 |
| MiMo-V2.5-Pro | 46.5 |
| Qwen3.6 27B | 31.2 |
| Qwen3.6 35B A3B | 24.7 |