OpenVLM VCR - Chinese Hard: leaderboard
VCR (Visual Caption Restoration) Chinese hard split from the OpenCompass OpenVLM Leaderboard (VLMEvalKit): the model restores Chinese caption text that is partly covered in the image; the easy split leaves more of each covered line visible than the hard split. Jaccard similarity between restored and reference words, in percent, higher is better.
Metric: Jaccard (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 50 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2 VL 7B | 43.8 | #743 |
| 2 | Qwen 2 VL 2B | 26.6 | #1191 |
| 3 | Gemini 1.5 Pro | 11.8 | #454 |
| 4 | InternVL2-8B | 10.3 | #813 |
| 5 | Claude 3 Opus | 9.2 | #565 |
| 6 | MiniCPM-V-2.6 | 7.3 | #827 |
| 7 | Claude 3.5 Sonnet | 4 | #362 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openvlm-vcr-chinese-hard · How It Works · Data refreshed daily, snapshot 2026-11-05.