OpenVLM Video - TempCompass - Caption Matching: leaderboard
TempCompass Caption Matching accuracy from the OpenCompass OpenVLM Video Leaderboard (VLMEvalKit): choosing the caption that matches the video on temporal aspects (action, speed, direction, order, attribute change) that a single frame cannot answer. Accuracy in percent, higher is better; an exact 0, a run with no scorable answers, is left out.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 35 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2 VL 72B | 86.36 | #450 |
| 2 | InternVL3-38B | 85.83 | #413 |
| 3 | InternVL3-78B | 84.5 | #361 |
| 4 | InternVL2.5-78B | 83.23 | #434 |
| 5 | Qwen 2.5 VL 7B | 80.71 | #641 |
| 6 | Qwen 2 VL 7B | 80.37 | #743 |
| 7 | Qwen2.5-Omni-7B | 79.51 | #633 |
| 8 | InternVL3-8B | 79.04 | #624 |
| 9 | Claude 3.5 Sonnet | 78.71 | #362 |
| 10 | InternVL2-8B | 76.38 | #813 |
| 11 | Gemini 1.5 Flash | 70.66 | #620 |
| 12 | MiniCPM-V-2.6 | 68.53 | #827 |
| 13 | Phi-4 Multimodal Instruct | 62.28 | #922 |
| 14 | Gemini 2.0 Flash | 16.03 | #351 |
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-tempcompass-caption-matching · How It Works · Data refreshed daily, snapshot 2026-11-05.