benchgap
Calibration

VideoMMMU → MMMU-Pro

MMMU-Pro is estimated from VideoMMMU with a Michaelis–Menten curve fitted on 13 models measured on both: y = 2.0000·x / (1.35131 + x), R² = 0.87, cross-validated error 3.5 pp. It is used for 0 estimates.

Fitted on

ModelVideoMMMUMMMU-Pro
Claude Opus 4.584.4%70.6%
dots3-note Preview86.8%79.1%
Gemini 3 Pro87.6%81.0%
GPT-5.285.9%79.5%
Kimi K2.586.6%78.5%
LFM3-VL-2B49.6%45.7%
MiniMax M384.6%78.1%
Qwen3.5 397B84.7%79.0%
Qwen3.6-27B84.4%75.8%
Qwen3.6-35B-A3B83.7%75.3%
Qwen3.6 Plus84.0%78.8%
Qwen3.7 Plus85.4%79.0%
Qwen3.8 Max88.7%82.3%