MMBench (English, test)
Multimodal · 2023-07-12
MMBench is a multiple-choice multimodal benchmark covering 20 perception and reasoning abilities, scored with CircularEval so a question counts only if answered correctly under every option rotation. This row is the original (v1.0) English test split.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen2.5-VL-72B-Instruct | 88.6 |
| InternVL2.5-78B | 88.3 |
| Qwen2.5-VL-7B-Instruct | 83.5 |
| GPT-4o (2024-05-13) | 83.4 |
| Qwen2 VL 7B Instruct | 83.0 |
| Qwen2.5-VL-3B-Instruct | 79.1 |
| Qwen2 VL 2B Instruct | 74.9 |