Multi-SWE-bench
Code · 2025-04-03
Multi-SWE-bench is ByteDance's multilingual issue-resolution benchmark of 1,632 human-validated GitHub issues across Java, TypeScript, JavaScript, Go, Rust, C and C++ repositories. An agent edits the repository snapshot and an instance counts as resolved when the fail-to-pass and pass-to-pass tests succeed. Scores report the resolved percentage; the agent harness is recorded on the config. Higher is better.
Top models (higher is better)
| Model | Score |
|---|---|
| Kimi K2 Thinking | 41.9 |
| Kimi K2 Instruct 0905 | 33.5 |
| Kimi K2 Instruct | 31.3 |