Calibration
SWE-bench Verified → NL2Repo
NL2Repo is estimated from SWE-bench Verified with a Hill curve fitted on 19 models measured on both: y = 0.0021 + (1.2000 − 0.0021)·x^6.00 / (0.86780^6.00 + x^6.00), R² = 0.65, cross-validated error 6.2 pp. It is used for 26 estimates.
| Estimated model | SWE-bench Verified | NL2Repo | Source |
|---|---|---|---|
| Ember-1 | 92.2% | 70.9% | estimated ± 6.2 pp, low confidence |
| BTL-4 | 78.4% | 42.4% | estimated ± 6.2 pp, medium confidence |
| MiMo-V2-Pro | 78.0% | 41.6% | estimated ± 6.2 pp, medium confidence |
| Claude Sonnet 4.5 | 77.2% | 39.9% | estimated ± 6.2 pp, medium confidence |
| MiMo-V2-Omni | 74.8% | 35.0% | estimated ± 6.2 pp, medium confidence |
| Claude 4.1 Opus | 74.5% | 34.5% | estimated ± 6.2 pp, medium confidence |
| Jais-3-MoVA-36B-A4B | 72.8% | 31.2% | estimated ± 6.2 pp, medium confidence |
| Claude 4 Sonnet | 72.7% | 31.0% | estimated ± 6.2 pp, medium confidence |
| MAI-Code-1.1-Flash | 72.6% | 30.8% | estimated ± 6.2 pp, medium confidence |
| Claude 4 Opus | 72.5% | 30.6% | estimated ± 6.2 pp, medium confidence |
| Jais-3-32B | 72.4% | 30.4% | estimated ± 6.2 pp, medium confidence |
| Grok Code Fast 1 | 70.8% | 27.5% | estimated ± 6.2 pp, medium confidence |
| Solar Pro 4 | 70.6% | 27.1% | estimated ± 6.2 pp, medium confidence |
| Jais-3-375B-A23B | 70.2% | 26.4% | estimated ± 6.2 pp, medium confidence |
| Jais-3-7B | 70.2% | 26.4% | estimated ± 6.2 pp, medium confidence |
| Qwen3-Coder 480B-A35B | 69.6% | 25.4% | estimated ± 6.2 pp, medium confidence |
| K-EXAONE 2.0 | 68.2% | 23.1% | estimated ± 6.2 pp, low confidence |
| DeepSeek V3.1 | 66.0% | 19.6% | estimated ± 6.2 pp, low confidence |
| Kimi K2 | 65.8% | 19.3% | estimated ± 6.2 pp, low confidence |
| Claude 3.7 Sonnet | 62.3% | 14.6% | estimated ± 6.2 pp, low confidence |
| Nemotron 3 Super 100B | 60.5% | 12.5% | estimated ± 6.2 pp, low confidence |
| GLM-4.7-Flash | 59.2% | 11.2% | estimated ± 6.2 pp, low confidence |
| DeepSeek-R1-0528 | 57.6% | 9.6% | estimated ± 6.2 pp, low confidence |
| MiniMax M1 80k | 56.0% | 8.3% | estimated ± 6.2 pp, low confidence |
| GPT-4.1 | 54.6% | 7.2% | estimated ± 6.2 pp, low confidence |
| ZAYA1-74B-Preview | 53.2% | 6.2% | estimated ± 6.2 pp, low confidence |
Fitted on
| Model | SWE-bench Verified | NL2Repo |
|---|---|---|
| Claude Opus 4.5 | 80.9% | 43.2% |
| DeepSeek V4 Flash 0731 | 79.0% | 54.2% |
| DeepSeek V4 Pro 0813 | 80.6% | 61.5% |
| dots3-note Preview | 78.4% | 49.8% |
| GLM-5 | 77.8% | 35.9% |
| Kimi K2.5 | 76.8% | 32.0% |
| MiniMax M3 | 80.5% | 42.1% |
| Ornith-1.0-35B | 75.6% | 34.6% |
| Ornith-1.0-397B | 82.4% | 48.2% |
| Ornith-1.0-9B | 69.4% | 27.2% |
| Ornith-1.5-35B-A3B | 79.0% | 46.2% |
| Ornith-1.5-397B | 86.0% | 59.5% |
| Ornith-1.5-9B | 70.6% | 32.4% |
| Qwen3.5 397B | 76.2% | 32.2% |
| Qwen3.6-27B | 77.2% | 36.2% |
| Qwen3.6-35B-A3B | 73.4% | 29.4% |
| Qwen3.6 Plus | 78.8% | 37.9% |
| Qwen3.7 Max | 80.4% | 47.2% |
| Qwen3.7 Plus | 77.7% | 41.1% |