MLVU (multiple-choice M-Avg)
Multimodal · 2024-06-06
MLVU is a long-video understanding benchmark with videos from about three minutes to over two hours across nine tasks. This row is M-Avg, the mean accuracy over its multiple-choice tasks, as release tables report it; most sources do not name the split (Qwen2.5-VL states the dev set), and the generative-task G-Avg is a separate metric. Higher is better.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen3.8-Max | 90.8 |
| GPT-5.6 Sol | 87.6 |
| Qwen3.7-Plus (source unspecified) | 87.4 |
| Qwen3.5 122B A10B | 87.3 |
| Qwen3.5-Omni-Plus | 86.8 |
| Qwen3.5 397B A17B | 86.7 |
| Qwen3.6 Plus (2026-04-02) | 86.7 |
| Qwen3.6 27B | 86.6 |
| Qwen3.6 35B A3B | 86.2 |
| GPT-5 | 86.2 |
| GPT-5.4 | 86.1 |
| Qwen3.5-27B | 85.9 |
| GPT-5.2 | 85.6 |
| Qwen3.5 35B A3B | 85.6 |
| Gemini 2.5 Pro | 85.6 |
| Qwen3.5 Plus (2026-02-15) | 85.1 |
| Kimi K2.5 | 85.0 |
| Gemini 3.1 Pro Preview | 84.7 |
| Qwen3.5-9B | 84.4 |
| Qwen3-VL-235B-A22B-Instruct | 84.3 |
| Qwen3 VL 235B A22B Thinking | 83.8 |
| GPT-5 Mini | 83.3 |
| Gemini 3 Pro Preview | 83.0 |
| Qwen3.5-4B | 82.8 |
| Qwen3 VL 32B Thinking | 82.3 |
| Gemini 2.5 Flash | 82.1 |
| Qwen3 VL 32B Instruct | 82.1 |
| Qwen3.5-Omni-Flash | 81.9 |
| Opus 4.5 | 81.7 |
| Opus 4.6 | 81.7 |
| Qwen3 VL 30B A3B Instruct | 81.3 |
| Qwen3 VL 30B A3B Thinking | 78.9 |
| Gemini 2.5 Flash-Lite | 78.5 |
| Gemini 3.5 Flash-Lite | 78.2 |
| InternVL3.5-241B-A28B | 78.2 |
| Qwen3-VL-8B-Instruct | 78.1 |
| GPT-5.6 Luna | 77.6 |
| InternVL2.5-78B | 75.7 |
| Qwen3 VL 4B Thinking | 75.7 |
| Qwen3 VL 4B Instruct | 75.3 |