OpenVLM Q-Bench: leaderboard
Q-Bench accuracy from the OpenCompass OpenVLM Leaderboard (VLMEvalKit): questions on low-level visual perception such as blur, noise, exposure and colour. Accuracy in percent, higher is better; exact-zero rows (failed runs) are dropped.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 160 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2 VL 72B | 80.3 | #450 |
| 2 | GPT-4.1 Mini | 80 | #363 |
| 3 | GPT-4.1 (2025-04-14) | 79.7 | #257 |
| 4 | InternVL3-78B | 78.8 | #361 |
| 5 | InternVL3-38B | 78.3 | #413 |
| 6 | Qwen 2.5 VL 7B | 77.7 | #641 |
| 7 | Moonshot v1 8K | 77.5 | #622 |
| 8 | Qwen 2 VL 7B | 76.9 | #743 |
| 9 | InternVL3-14B | 76.7 | #506 |
| 10 | MiniCPM-V-2.6 | 76.5 | #827 |
| 11 | InternVL3-8B | 75.9 | #624 |
| 12 | GPT-4.1 Nano | 74.8 | #730 |
| 13 | Gemini 1.5 Pro | 74.7 | #454 |
| 14 | Aquila-VL-2B | 74.2 | #1216 |
| 15 | Qwen 2 VL 2B | 73.9 | #1191 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openvlm-q-bench · How It Works · Data refreshed daily, snapshot 2026-11-05.