OpenVLM Video - Video-MME (w/o subs) - Counting Problem: leaderboard
Video-MME Counting Problem accuracy without subtitles, from the OpenCompass OpenVLM Video Leaderboard (VLMEvalKit): the four-option multiple-choice questions of this task type, pooled over short, medium and long videos. One of 12 Video-MME task types; 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 | Gemini 2.0 Flash | 48.9 | #351 |
| 2 | InternVL3-78B | 48.5 | #361 |
| 3 | InternVL2.5-78B | 48.5 | #434 |
| 4 | InternVL3-38B | 45.9 | #413 |
| 5 | InternVL3-8B | 44 | #624 |
| 6 | Qwen 2 VL 72B | 41.4 | #450 |
| 7 | Qwen 2.5 VL 7B | 41 | #641 |
| 8 | Qwen2.5-Omni-7B | 40.3 | #633 |
| 9 | MiniCPM-V-2.6 | 39.2 | #827 |
| 10 | Gemini 1.5 Flash | 38.1 | #620 |
| 11 | InternVL2-8B | 35.8 | #813 |
| 12 | Claude 3.5 Sonnet | 35.4 | #362 |
| 13 | Qwen 2 VL 7B | 34.7 | #743 |
| 14 | Phi-4 Multimodal Instruct | 28 | #922 |
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-video-mme-w-o-subs-counting-problem · How It Works · Data refreshed daily, snapshot 2026-11-05.