OpenVLM A-Bench - Technical Quality: leaderboard
A-Bench Technical Quality accuracy from the OpenCompass OpenVLM Leaderboard (VLMEvalKit): multiple-choice questions on AI-generated images, judging technical quality such as blur, noise and exposure. A dimension of A-Bench-P2 (low-level quality perception). 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 | MiniCPM-V-2.6 | 77.2 | #827 |
| 2 | Qwen 2 VL 72B | 76.4 | #450 |
| 3 | InternVL3-78B | 75.3 | #361 |
| 4 | GPT-4.1 (2025-04-14) | 75.3 | #257 |
| 5 | InternVL3-14B | 73.7 | #506 |
| 6 | Grok 2 (1212) | 73.4 | #502 |
| 7 | GPT-4.1 Mini | 72.2 | #363 |
| 8 | Gemini 1.5 Flash | 72.2 | #620 |
| 9 | InternVL3-38B | 71.8 | #413 |
| 10 | Aquila-VL-2B | 71.8 | #1216 |
| 11 | Qwen 2.5 VL 7B | 71.8 | #641 |
| 12 | InternVL3-8B | 71 | #624 |
| 13 | Qwen 2 VL 7B | 71 | #743 |
| 14 | Gemma 3 27B | 71 | #606 |
| 15 | Pixtral-12B | 71 | #805 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openvlm-a-bench-technical-quality · How It Works · Data refreshed daily, snapshot 2026-11-05.