VideoMME w/o sub.: leaderboard

Frames-only setting of Video-MME (2024): 2,700 four-option questions on 900 videos of 11 s to 1 h, 300 each of short, medium (4 to 15 min) and long (30 to 60 min); Gemini 1.5 Pro 75%, GPT-4o 71.9%.

Metric: Score (self-reported). Source: arxiv.org. 51 models tracked.

Top models

#ModelScore
1video-SALMONN 2+79.7
2Gemini 1.5 Pro75
3AdaReTaKe73.5
4TSPO73
5Keye-VL-1.573
6InternVL2.572.1
7GPT-4o71.9
8Qwen2-VL71.2
9LLaVA-Video70.6
10JT-VL-Chat70.6

Interactive version: theaggregate.ai/benchmark?slug=videomme-w-o-sub · How It Works · Data refreshed daily, snapshot 2026-09-05.