benchgap
Calibration

SciCode → Vals SWE-bench

Vals SWE-bench is estimated from SciCode with a Hill curve fitted on 8 models measured on both: y = 0.0000 + (0.7966 − 0.0000)·x^6.00 / (0.31757^6.00 + x^6.00), R² = 0.39, cross-validated error 5.3 pp. It is used for 19 estimates.

Estimated modelSciCodeVals SWE-benchSource
A.X K241.0%65.5%estimated ± 5.3 pp, low confidence
Agents-A1-4B29.6%31.6%estimated ± 5.3 pp, low confidence
Granite 4.2 30B38.8%61.2%estimated ± 5.3 pp, low confidence
Granite 4.2 3B24.1%12.8%estimated ± 5.3 pp, low confidence
Granite 4.2 8B36.1%54.4%estimated ± 5.3 pp, low confidence
Hy3 Preview41.2%65.9%estimated ± 5.3 pp, low confidence
K-EXAONE 2.037.4%57.9%estimated ± 5.3 pp, low confidence
Ling 2.6 Flash27.0%21.8%estimated ± 5.3 pp, low confidence
Ling 3.0 Flash FP840.4%64.4%estimated ± 5.3 pp, low confidence
Mercury 2.538.0%59.4%estimated ± 5.3 pp, low confidence
MiniCPM5-2B26.3%19.4%estimated ± 5.3 pp, low confidence
Muse Glimmer 30B43.6%69.3%estimated ± 5.3 pp, low confidence
Nemotron 3.5 Lightning 30B A3B NVFP431.4%38.4%estimated ± 5.3 pp, low confidence
Nemotron 3 Nano Omni 30B A3B32.0%40.7%estimated ± 5.3 pp, low confidence
Qwen3.7 Plus51.3%75.4%estimated ± 5.3 pp, low confidence
Beam49.7%74.6%estimated ± 5.3 pp, low confidence
Sakana Fugu60.1%78.0%estimated ± 5.3 pp, low confidence
Sakana Fugu-Ultra58.7%77.7%estimated ± 5.3 pp, low confidence
Step 5 Preview58.9%77.8%estimated ± 5.3 pp, low confidence