74.1%
MVBench average accuracy of JT-VL-Chat and JT3.5, tied first in early 2026; Woven-VLLM 73.8%
What the chart shows
MVBench tests whether multimodal models can follow motion, sequences and changing context across video frames rather than in single images. In early 2026 JT3.5 and JT-VL-Chat tied at the top with 74.1% average accuracy, just ahead of Woven-VLLM (73.8%).
The top 15 span 23 points, down to VideoChat2 at 51.1%. Later variants of the same model score higher: VideoChat2_mistral reached 60.4% and VideoChat2_HD_mistral 62.3%.
Key numbers
| Measure | Value |
|---|---|
| JT3.5 | 74.075% |
| JT-VL-Chat | 74.075% |
| Woven-VLLM | 73.81% |
| Video-CCAM-7B-v1.2 | 69.23% |
| VideoChat2 (15th) | 51.1% |
How to read it
- Snapshot of the MVBench leaderboard (OpenGVLab) in early 2026, top 15 models only.
- Accuracy is averaged across the benchmark’s multiple-choice video tasks.
- Several entries are variants of the same model family (VideoChat2, Video-CCAM).
Source: MVBench Leaderboard, 2026, via the AI Index Report 2026 (Stanford HAI), figure 2.3.1: MVBench: average accuracy. Chart and text by AI Adoption.
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Discussion