MLVU-Dev: leaderboard
MLVU (BAAI, CVPR 2025) tests multimodal models on long videos of 3 minutes to 2 hours. This board is the Dev set's multiple-choice average (M-Avg) over topic reasoning, anomaly recognition, needle QA, ego reasoning, plot QA, action order and action count, as accuracy in percent; higher is better. Each entry samples frames at its own rate or count.
Metric: Multiple-choice accuracy (%). Source: github.com. Saturation forecast: Estimated already saturated. 39 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 64.6 | #741 |
| 2 | GPT-4 Turbo | 49.2 | #1002 |
| 3 | Qwen-VL-Max | 42.2 | #1226 |
| 4 | Claude 3 Opus | 36.5 | #1459 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mlvu-dev · How It Works · Data refreshed daily, snapshot 2026-11-12.