Calibration
SciCode → SWE-bench Pro
SWE-bench Pro is estimated from SciCode with a Hill curve fitted on 19 models measured on both: y = 0.0379 + (0.6771 − 0.0379)·x^6.00 / (0.35687^6.00 + x^6.00), R² = 0.71, cross-validated error 9.7 pp. It is used for 1 estimate.
| Estimated model | SciCode | SWE-bench Pro | Source |
|---|---|---|---|
| Ling 3.0 Flash FP8 | 40.4% | 47.1% | estimated ± 9.7 pp, medium confidence |
Fitted on
| Model | SciCode | SWE-bench Pro |
|---|---|---|
| Gemini 3.5 Flash | 53.1% | 55.1% |
| Granite 4.2 30B | 38.8% | 33.3% |
| Granite 4.2 8B | 36.1% | 19.1% |
| Inkling-Small | 48.7% | 55.9% |
| Kimi K2.6 | 52.2% | 58.6% |
| Kimi K2.5 | 48.7% | 50.7% |
| Ling 3.0 Flash | 41.2% | 56.6% |
| MiniCPM5-2B | 26.3% | 14.4% |
| Muse Glimmer 30B | 43.6% | 51.2% |
| Qwen 3.6 Max (preview) | 47.0% | 57.3% |
| Qwen3.7 Max | 53.5% | 60.6% |
| Qwen3.7 Plus | 51.3% | 57.6% |
| Beam 1.0 501B-A23B | 49.2% | 66.1% |
| Beam 1.0 501B-A23B FP8 | 48.3% | 65.9% |
| Beam 1.0 501B-A23B NVFP4 | 46.9% | 65.6% |
| Sage-A1.5 | 44.3% | 61.2% |
| Sakana Fugu | 60.1% | 59.0% |
| Sakana Fugu-Ultra | 58.7% | 73.7% |
| Spark-X2.6-4B | 36.8% | 54.3% |