benchgap
Calibration

LiveCodeBench v6 → SWE-bench Verified

SWE-bench Verified is estimated from LiveCodeBench v6 with a Michaelis–Menten curve fitted on 15 models measured on both: y = 2.0000·x / (1.55881 + x), R² = 0.32, cross-validated error 10.5 pp. It is used for 6 estimates.

Estimated modelLiveCodeBench v6SWE-bench VerifiedSource
BTL-388.1%72.2%estimated ± 10.5 pp, low confidence
LLaDA2.2-mini28.1%30.6%estimated ± 10.5 pp, low confidence
Mellum2-12B-A2.5B-Instruct37.2%38.5%estimated ± 10.5 pp, low confidence
Mellum2-12B-A2.5B-Thinking69.9%61.9%estimated ± 10.5 pp, low confidence
MiniCPM5-1B33.5%35.4%estimated ± 10.5 pp, low confidence
ZAYA1-8B65.8%59.4%estimated ± 10.5 pp, low confidence