benchgap
Calibration

Video-MME (with subtitle) → CountBench

CountBench is estimated from Video-MME (with subtitle) with a offset logistic curve fitted on 7 models measured on both: y = 0.9164 + (0.9839 − 0.9164) / (1 + exp(−200.00·(x − 0.8717))), R² = 0.83, cross-validated error 2.0 pp. It is used for 2 estimates.

Estimated modelVideo-MME (with subtitle)CountBenchSource
MiMo-V2.587.7%96.7%estimated ± 2.0 pp, medium confidence
MiniMax M385.4%91.8%estimated ± 2.0 pp, medium confidence

Fitted on

ModelVideo-MME (with subtitle)CountBench
Claude Opus 4.577.6%90.6%
Gemini 3 Pro88.4%97.3%
GPT-5.286.0%91.9%
Kimi K2.587.4%94.1%
LFM3-VL-2B69.0%93.0%
Qwen3.5 397B87.5%97.2%
Qwen3.6-27B87.7%97.8%