OpenVLM Video - TempCompass - Action: leaderboard
TempCompass Action accuracy from the OpenCompass OpenVLM Video Leaderboard (VLMEvalKit): questions on which action is performed in videos where a single frame cannot answer, pooled over the multi-choice, yes-or-no, caption-matching and captioning formats. Percent, higher is better; runs with a failed format (exactly 0) are left out.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 34 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2 VL 72B | 97.1 | #450 |
| 2 | InternVL3-38B | 95.75 | #413 |
| 3 | Qwen 2.5 VL 7B | 94.64 | #641 |
| 4 | InternVL3-78B | 93.84 | #361 |
| 5 | Qwen2.5-Omni-7B | 93.84 | #633 |
| 6 | InternVL3-8B | 93.35 | #624 |
| 7 | Qwen 2 VL 7B | 93.22 | #743 |
| 8 | Gemini 1.5 Flash | 92.61 | #620 |
| 9 | Claude 3.5 Sonnet | 89.53 | #362 |
| 10 | InternVL2-8B | 89.22 | #813 |
| 11 | MiniCPM-V-2.6 | 84.66 | #827 |
| 12 | Phi-4 Multimodal Instruct | 81.7 | #922 |
| 13 | Gemini 2.0 Flash | 78 | #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-action · How It Works · Data refreshed daily, snapshot 2026-11-05.