OpenVLM A-Bench: leaderboard
OpenCompass OpenVLM evaluation of A-Bench: attribute and compositional image understanding with questions about object attributes, spatial arrangement, occlusion, orientation, size, and relations.
Metric: Accuracy (%). Source: huggingface.co. Status: saturated. 160 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-4.1 (2025-04-14) | 79.6 |
| 2 | Qwen 2 VL 72B | 79.4 |
| 3 | InternVL3-78B | 78.9 |
| 4 | GPT-4.1 Mini | 77.4 |
| 5 | InternVL3-38B | 77.1 |
| 6 | Qwen 2 VL 7B | 76.4 |
| 7 | InternVL3-8B | 76.3 |
| 8 | Gemma 3 27B | 76.3 |
| 9 | Grok 2 (1212) | 76.3 |
| 10 | Aquila-VL-2B | 75.4 |
| 11 | Gemini 1.5 Pro | 75.4 |
| 12 | Gemini 1.5 Flash | 73.6 |
| 13 | Pixtral-12B | 73.1 |
| 14 | InternVL2-8B | 72.9 |
| 15 | GPT-4.1 Nano | 72.9 |
Interactive version: theaggregate.ai/benchmark?slug=openvlm-a-bench · How It Works · Data refreshed daily, snapshot 2026-09-05.