benchgap
Calibration

SciCode → SWE-bench Verified

SWE-bench Verified is estimated from SciCode with a Hill curve fitted on 19 models measured on both: y = 0.3879 + (0.8864 − 0.3879)·x^6.00 / (0.38521^6.00 + x^6.00), R² = 0.90, cross-validated error 5.1 pp. It is used for 7 estimates.

Estimated modelSciCodeSWE-bench VerifiedSource
A.X K241.0%68.3%estimated ± 5.1 pp, medium confidence
Ling 3.0 Flash FP840.4%67.2%estimated ± 5.1 pp, medium confidence
Mercury 2.538.0%62.7%estimated ± 5.1 pp, medium confidence
Nemotron 3 Nano Omni 30B A3B32.0%51.1%estimated ± 5.1 pp, medium confidence
Agents-A1-4B29.6%47.3%estimated ± 5.1 pp, medium confidence
Ling 2.6 Flash27.0%44.1%estimated ± 5.1 pp, medium confidence
Granite 4.2 3B24.1%41.6%estimated ± 5.1 pp, low confidence

Fitted on

ModelSciCodeSWE-bench Verified
Granite 4.2 30B38.8%57.0%
Granite 4.2 8B36.1%47.7%
Hy3 Preview41.2%74.4%
Inkling-Small48.7%80.2%
K-EXAONE 2.037.4%68.2%
Kimi K2.652.2%80.2%
Kimi K2.548.7%76.8%
LFM2.5-24B-A2B32.7%51.0%
LFM3-1.7B28.5%43.4%
MiniCPM5-2B26.3%46.4%
Muse Glimmer 30B43.6%76.0%
Nemotron 3.5 Lightning 30B A3B NVFP431.4%52.8%
Nemotron 3 Ultra44.6%71.9%
Qwen3.7 Max53.5%80.4%
Qwen3.7 Plus51.3%77.7%
Beam 1.0 501B-A23B49.2%81.6%
Beam 1.0 501B-A23B FP848.3%80.2%
Beam 1.0 501B-A23B NVFP446.9%80.7%
Spark-X2.6-4B36.8%62.8%