benchgap
Calibration

Video-MME (with subtitle) → VideoMMMU

VideoMMMU is estimated from Video-MME (with subtitle) with a inverse Michaelis–Menten curve fitted on 11 models measured on both: y = 1.25047·(x − 0.0579) / (0.0579 + 2.0000 − x), R² = 0.75, cross-validated error 6.5 pp. It is used for 0 estimates.

Fitted on

ModelVideo-MME (with subtitle)VideoMMMU
Claude Opus 4.577.6%84.4%
Gemini 3 Pro88.4%87.6%
GPT-5.286.0%85.9%
Kimi K2.587.4%86.6%
LFM3-VL-2B69.0%49.6%
MiniMax M385.4%84.6%
Qwen3.5 397B87.5%84.7%
Qwen3.6-27B87.7%84.4%
Qwen3.6-35B-A3B86.6%83.7%
Qwen3.7 Plus88.0%85.4%
Qwen3.8 Max90.4%88.7%