OpenVLM Video - MVBench - Scene Transition: leaderboard

MVBench Scene Transition accuracy from the OpenCompass OpenVLM Video Leaderboard (VLMEvalKit): one of MVBench's 20 temporal tasks, 200 multiple-choice questions on how the scene changes over the clip. Accuracy in percent, higher is better.

Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 37 models tracked.

Top models

#ModelScoreOverall rank
1InternVL3-78B97#361
2Qwen 2 VL 72B95#450
3Qwen 2 VL 7B94.5#743
4InternVL3-8B93#624
5Gemini 2.0 Flash93#351
6InternVL3-38B93#413
7Qwen 2.5 VL 7B92.5#641
8InternVL2.5-78B92.5#434
9Qwen2.5-Omni-7B92.5#633
10InternVL2-8B88.5#813
11Gemini 1.5 Flash84.5#620
12MiniCPM-V-2.677.5#827
13Phi-4 Multimodal Instruct77.5#922
14Claude 3.5 Sonnet42.5#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-mvbench-scene-transition · How It Works · Data refreshed daily, snapshot 2026-11-05.