OpenVLM Video - MLVU - Generation Score: leaderboard
MLVU generation score from the OpenCompass OpenVLM Video Leaderboard (VLMEvalKit): the mean over sub-scene captioning and video summarization items of two 1-5 LLM-judge ratings added together (accuracy and relevance, or completeness and reliability), so 2-10 per scored item. Higher is better; a run with no scored item (0) is left out.
Metric: Judge Score (0-10). Source: huggingface.co. Saturation forecast: Estimated already saturated. 36 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.0 Flash | 4.9 | #351 |
| 2 | InternVL3-38B | 4.8 | #413 |
| 3 | InternVL3-78B | 4.5 | #361 |
| 4 | Qwen 2.5 VL 7B | 4.5 | #641 |
| 5 | InternVL2.5-78B | 4.5 | #434 |
| 6 | MiniCPM-V-2.6 | 4.45 | #827 |
| 7 | Qwen2.5-Omni-7B | 4.3 | #633 |
| 8 | Qwen 2 VL 72B | 4.2 | #450 |
| 9 | InternVL3-8B | 4.1 | #624 |
| 10 | Qwen 2 VL 7B | 4.1 | #743 |
| 11 | Claude 3.5 Sonnet | 3.89 | #362 |
| 12 | InternVL2-8B | 3.84 | #813 |
| 13 | Gemini 1.5 Flash | 3.71 | #620 |
| 14 | Phi-4 Multimodal Instruct | 3.3 | #922 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openvlm-video-mlvu-generation-score · How It Works · Data refreshed daily, snapshot 2026-11-05.