benchgap
Calibration

Video-MME (w/o subtitle) → MMMU

MMMU is estimated from Video-MME (w/o subtitle) with a Hill curve fitted on 8 models measured on both: y = 0.0000 + (1.1348 − 0.0000)·x^3.77 / (0.63364^3.77 + x^3.77), R² = 0.99, cross-validated error 2.0 pp. It is used for 0 estimates.

Fitted on

ModelVideo-MME (w/o subtitle)MMMU
Claude Opus 4.581.4%80.7%
Gemini 3 Pro87.7%87.2%
GPT-5.285.8%86.7%
Kimi K2.583.2%84.3%
LFM3-VL-2B62.8%55.7%
Nemotron 3 Nano Omni 30B A3B72.2%70.8%
Qwen3.5 397B83.7%85.0%
Qwen3.6-35B-A3B82.5%81.7%