OpenVLM MMBench V1.0 CN - AR: leaderboard
MMBench v1.0 Chinese test questions on attribute reasoning (AR: physical property, function and identity reasoning), from the OpenCompass OpenVLM Leaderboard (VLMEvalKit). CircularEval counts a question only if every rotation of its options is answered correctly. Accuracy in percent, higher is better; this is the original question set, separate from the MMBench V1.1 boards.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 230 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | InternVL3-38B | 88.5 | #413 |
| 2 | InternVL3-78B | 87.8 | #361 |
| 3 | GPT-4.1 (2025-04-14) | 87.2 | #257 |
| 4 | InternVL3-14B | 85.4 | #506 |
| 5 | InternVL2.5-78B | 85.4 | #434 |
| 6 | Grok 2 (1212) | 85.4 | #502 |
| 7 | Qwen 2 VL 72B | 85.1 | #450 |
| 8 | GPT-4.1 Mini | 84.4 | #363 |
| 9 | InternVL3-8B | 84 | #624 |
| 10 | Gemma 3 27B | 83.3 | #606 |
| 11 | Qwen 2.5 VL 7B | 83.3 | #641 |
| 12 | Aquila-VL-2B | 83 | #1216 |
| 13 | MiniCPM-V-2.6 | 82.6 | #827 |
| 14 | Qwen 2 VL 7B | 82.6 | #743 |
| 15 | Claude 3.5 Sonnet | 82.6 | #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-mmbench-v1-0-cn-ar · How It Works · Data refreshed daily, snapshot 2026-11-05.