OpenVLM A-Bench - Object Counting: leaderboard
A-Bench Object Counting accuracy from the OpenCompass OpenVLM Leaderboard (VLMEvalKit): multiple-choice questions on AI-generated images, counting objects. A dimension of A-Bench-P1 (high-level semantic understanding). 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 Mini | 89.8 | #363 |
| 2 | InternVL3-14B | 85.7 | #506 |
| 3 | Claude 3.7 Sonnet | 85.7 | #259 |
| 4 | Qwen 2 VL 72B | 85.7 | #450 |
| 5 | Qwen 2.5 VL 7B | 83.7 | #641 |
| 6 | InternVL3-38B | 81.6 | #413 |
| 7 | InternVL3-78B | 81.6 | #361 |
| 8 | GPT-4.1 (2025-04-14) | 81.6 | #257 |
| 9 | Qwen 2 VL 7B | 79.6 | #743 |
| 10 | Pixtral-12B | 79.6 | #805 |
| 11 | Grok 2 (1212) | 79.6 | #502 |
| 12 | Qwen 2 VL 2B | 77.6 | #1191 |
| 13 | Llama 3.2 11B Instruct | 77.6 | #1108 |
| 14 | Gemini 1.5 Flash | 77.6 | #620 |
| 15 | InternVL3-8B | 75.5 | #624 |
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-object-counting · How It Works · Data refreshed daily, snapshot 2026-11-05.