MMBench v1.1 (dev, English)
Multimodal · 2023-07-12
MMBench v1.1 is the refined release of OpenCompass's multiple-choice benchmark of perception and reasoning across 20 ability dimensions. This row is the English dev split, scored with CircularEval: a question counts only when the model answers every rotated option order correctly. The English test split and the v1.0 splits are separate benchmarks. Higher is better.
Top models (higher is better)
| Model | Score |
|---|---|
| Kimi K2.5 | 94.2 |
| Qwen3.5 Plus (2026-02-15) | 93.8 |
| Gemini 3 Pro Preview | 93.7 |
| Qwen3.5 397B A17B | 93.7 |
| Qwen3.5 122B A10B | 92.8 |
| Qwen3.5-Omni-Plus | 92.8 |
| Qwen3.6 35B A3B | 92.8 |
| Qwen3.5-27B | 92.6 |
| Qwen3.6 27B | 92.3 |
| Qwen3.5 35B A3B | 91.5 |
| Gemma 4 31B IT | 90.9 |
| Qwen3 VL 32B Thinking | 90.8 |
| Qwen3 VL 235B A22B Thinking | 90.6 |
| Qwen3.5-9B | 90.1 |
| Gemini 2.5 Pro | 90.1 |
| Qwen3-VL-235B-A22B-Instruct | 89.9 |
| Qwen3.5-4B | 89.4 |
| Opus 4.5 | 89.2 |
| Gemma 4 26B A4B IT | 89.0 |
| Qwen3 VL 30B A3B Thinking | 88.9 |
| Qwen3 VL 32B Instruct | 88.9 |
| Qwen3.5-Omni-Flash | 88.8 |
| Qwen2.5-VL-72B-Instruct | 88.6 |
| Sonnet 4.5 | 88.3 |
| Gemini 3.5 Flash-Lite | 88.2 |
| Gemma 4 12B IT | 88.2 |
| GPT-5.2 | 88.2 |
| Qwen3 VL 8B Thinking | 87.5 |
| Qwen3 VL 30B A3B Instruct | 87.0 |
| GPT-5 Mini | 86.8 |
| Qwen3 VL 4B Thinking | 86.7 |
| LFM3-VL-2B | 86.1 |
| GPT-5 | 85.9 |
| Qwen3 VL 4B Instruct | 85.1 |
| Qwen3-VL-8B-Instruct | 85.0 |
| Opus 4.1 | 84.9 |
| GPT-5.6 Luna | 84.4 |
| Gemini 2.5 Flash-Lite | 82.7 |
| Sonnet 4 | 82.4 |
| Qwen3 VL 2B Thinking | 81.9 |