benchgap
Calibration

VITA-Bench → QwenClawBench

QwenClawBench is estimated from VITA-Bench with a offset logistic curve fitted on 6 models measured on both: y = 0.5204 + (0.6396 − 0.5204) / (1 + exp(−200.00·(x − 0.4465))), R² = 0.96, cross-validated error 2.0 pp. It is used for 7 estimates.

Estimated modelVITA-BenchQwenClawBenchSource
Sage-A1.545.4%61.8%estimated ± 2.0 pp, medium confidence
Agents-A1-4B40.3%52.0%estimated ± 2.0 pp, medium confidence
Agents-A138.8%52.0%estimated ± 2.0 pp, medium confidence
Claude Sonnet 4.517.0%52.0%estimated ± 2.0 pp, low confidence
DeepSeek V3.218.5%52.0%estimated ± 2.0 pp, low confidence
GLM-4.715.5%52.0%estimated ± 2.0 pp, low confidence
LongCat-Flash-Lite-Sparse21.7%52.0%estimated ± 2.0 pp, low confidence

Fitted on

ModelVITA-BenchQwenClawBench
Claude Opus 4.523.3%52.3%
Qwen3.5 397B43.7%51.8%
Qwen3.6-35B-A3B35.6%52.6%
Qwen3.6 Plus44.3%57.2%
Qwen3.7 Max47.9%64.3%
Qwen3.7 Plus45.6%61.8%