Benchmarks · coding
SWE-bench Multilingual
A 300-instance SWE-bench variant with tasks from 42 repositories across 9 programming languages. Maintained by SWE-bench Team. Score: % Resolved: percentage of task instances solved.
25 models · 25 results · data as of 2026-10-29 · swebench.com ↗
Leaderboard
Opus 5 max
All results
25 rows| # | Model | Effort | SWE-bench Multilingual | Harness |
|---|---|---|---|---|
| 1 | Opus 5 | max | 89.5% | |
| 2 | Fable 5.1 | max | 89.1% | |
| 3 | Fable 5 | max | 86.6% | |
| 4 | Opus 4.8 | max | 84.4% | |
| 5 | Qwen 3.8 Flash | max | 81.0% | mini-SWE-agent harness |
| 6 | Opus 4.7 | max | 80.5% | |
| 7 | Kimi K2.6 | max | 76.7% | in-house evaluation framework adapted from SWE-agent (minimal tools: bash, createfile, insert, view, strreplace, submit) |
| 8 | Qwen 3.7 Plus | high | 75.8% | mini-SWE-agent harness |
| 9 | Qwen 3.8 27B | max | 73.8% | mini-SWE-agent harness |
| 10 | Gemini 3 Flash | high | 72.7% | mini-SWE-agent |
| 11 | Opus 4.6 | max | 72.0% | mini-SWE-agent |
| 12 | MiMo V2 Flash | high | 71.7% | |
| 13 | Qwen 3.6 27B | max | 71.3% | internal agent scaffold (bash + file-edit tools) |
| 14 | Opus 4.5 | high | 70.7% | mini-SWE-agent |
| 15 | GLM 5 | high | 69.7% | mini-SWE-agent |
| 16 | Qwen 3.5 27B | high | 69.3% | internal agent scaffold (bash + file-edit tools) |
| 17 | Gemini 3 Pro | high | 68.7% | mini-SWE-agent |
| 18 | MiniMax M2.5 | high | 68.3% | mini-SWE-agent |
| 19 | Kimi K2.5 | max | 67.3% | in-house evaluation framework adapted from SWE-agent (minimal tools: bash, createfile, insert, view, strreplace, submit) |
| 20 | Qwen 3.6 35B-A3B | max | 67.2% | internal agent scaffold (bash + file-edit tools) |
| 21 | Sonnet 4.5 | high | 67.0% | mini-SWE-agent |
| 22 | GPT-5.2 | high | 66.7% | mini-SWE-agent |
| 23 | GPT-5.2 Codex | xhigh | 66.3% | mini-SWE-agent |
| 24 | DeepSeek V3.2 | high | 59.0% | mini-SWE-agent |
| 25 | GPT-5 Mini | high | 39.7% | mini-SWE-agent |