OpenVLM Q-Bench - How (In-context Distortion): leaderboard
Q-Bench accuracy on how questions about distortions of a particular object or region, from the OpenCompass OpenVLM Leaderboard (VLMEvalKit) on the Q-Bench dev split. Q-Bench probes low-level visual perception; each of its twelve question-type by concern cells holds 70-188 of the 1,495 dev questions. Accuracy in percent, higher is better.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 160 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 (2025-04-14) | 75.9 | #257 |
| 2 | Qwen 2 VL 72B | 71.3 | #450 |
| 3 | InternVL3-78B | 70.1 | #361 |
| 4 | GPT-4.1 Mini | 69 | #363 |
| 5 | InternVL3-38B | 69 | #413 |
| 6 | Gemini 1.5 Pro | 69 | #454 |
| 7 | MiniCPM-V-2.6 | 67.8 | #827 |
| 8 | Qwen 2 VL 7B | 66.7 | #743 |
| 9 | Qwen 2.5 VL 7B | 66.7 | #641 |
| 10 | Grok 2 (1212) | 66.7 | #502 |
| 11 | Pixtral-12B | 65.5 | #805 |
| 12 | Aquila-VL-2B | 65.5 | #1216 |
| 13 | Qwen 2 VL 2B | 64.4 | #1191 |
| 14 | InternVL3-8B | 63.2 | #624 |
| 15 | Gemini 1.5 Flash | 63.2 | #620 |
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-in-context-distortion · How It Works · Data refreshed daily, snapshot 2026-11-05.