OpenVLM Video - MMBench-Video - Hallucination: leaderboard
MMBench-Video Hallucination (HL) score from the OpenCompass OpenVLM Video Leaderboard (VLMEvalKit): open-ended answers to questions on whether the answer avoids content that is not in the video, rated by GPT-4 against a reference from 0 to 3 and averaged; higher is better. For models listed at several frame budgets (Gemini-1.5-Pro, GPT-4o, Aria) the 1 fps run is used.
Metric: GPT-4 Rating (0-3). Source: huggingface.co. Saturation forecast: Estimated already saturated. 45 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.0 Flash | 2.24 | #351 |
| 2 | Claude 3.5 Sonnet | 1.4 | #362 |
| 3 | Gemini 1.5 Flash | 1.29 | #620 |
| 4 | Qwen 2.5 VL 7B | 1.24 | #641 |
| 5 | MiniCPM-V-2.6 | 1.05 | #827 |
| 6 | InternVL2.5-78B | 0.89 | #434 |
| 7 | Qwen 2 VL 72B | 0.77 | #450 |
| 8 | Qwen2.5-Omni-7B | 0.74 | #633 |
| 9 | Phi-4 Multimodal Instruct | 0.71 | #922 |
| 10 | Qwen 2 VL 7B | 0.58 | #743 |
| 11 | InternVL3-78B | 0.35 | #361 |
| 12 | InternVL3-38B | 0.29 | #413 |
| 13 | InternVL3-8B | 0.26 | #624 |
| 14 | InternVL2-8B | 0.19 | #813 |
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-mmbench-video-hallucination · How It Works · Data refreshed daily, snapshot 2026-11-05.