OpenVLM Video - MLVU - Needle QA: leaderboard
MLVU Needle QA accuracy from the OpenCompass OpenVLM Video Leaderboard (VLMEvalKit): multiple-choice questions on long videos (3 minutes to 2 hours) about a short clip inserted into an unrelated long video. One of the seven MLVU multiple-choice tasks; accuracy in percent, higher is better.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 37 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | InternVL3-78B | 84.2 | #361 |
| 2 | InternVL3-38B | 80.8 | #413 |
| 3 | Qwen 2 VL 72B | 78.9 | #450 |
| 4 | InternVL2.5-78B | 78.6 | #434 |
| 5 | InternVL3-8B | 77.7 | #624 |
| 6 | Qwen 2 VL 7B | 76.1 | #743 |
| 7 | Qwen2.5-Omni-7B | 74.9 | #633 |
| 8 | Gemini 2.0 Flash | 72.4 | #351 |
| 9 | Qwen 2.5 VL 7B | 71.5 | #641 |
| 10 | MiniCPM-V-2.6 | 63.7 | #827 |
| 11 | InternVL2-8B | 60.8 | #813 |
| 12 | Phi-4 Multimodal Instruct | 56.1 | #922 |
| 13 | Gemini 1.5 Flash | 47 | #620 |
| 14 | Claude 3.5 Sonnet | 20.8 | #362 |
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-needle-qa · How It Works · Data refreshed daily, snapshot 2026-11-05.