Benchmarks · coding
NL2Repo-Bench
Build a complete, installable Python repository from a natural-language spec and an empty workspace (104 tasks, graded by the upstream pytest suites). Average test pass rate; scores are vendor and paper reports, gathered by LLMLearner.
18 models · 18 results · data as of 2026-10-29 · llmlearner.com ↗
Leaderboard
Opus 5 max
All results
18 rows| # | Model | Effort | NL2Repo-Bench |
|---|---|---|---|
| 1 | Opus 5 | max | 75.3% |
| 2 | DeepSeek V4.1 Flash 1017 | max | 68.2% |
| 3 | DeepSeek V4.1 Flash | max | 65.4% |
| 4 | Qwen 3.8 Max | xhigh | 64.9% |
| 5 | DeepSeek V4 Pro 0813 | max | 61.5% |
| 6 | MiniMax M3.1 | max | 61.2% |
| 7 | Hy4 | high | 58.9% |
| 8 | GLM 5.3 | max | 58.0% |
| 9 | Kimi K3 | max | 58.0% |
| 10 | DeepSeek V4 Flash | max | 57.7% |
| 11 | GPT-5.6 Sol | max | 56.8% |
| 12 | GLM 5.3 Flash | max | 56.3% |
| 13 | DeepSeek V4 Flash 0731 | max | 54.2% |
| 14 | GLM 5.2 | max | 48.9% |
| 15 | Qwen 3.8 Flash | xhigh | 48.1% |
| 16 | Qwen 3.8 27B | max | 42.3% |
| 17 | MiniMax M2.7 | high | 39.8% |
| 18 | DeepSeek V4 Pro | xhigh | 38.5% |