MMStar
Multimodal · 2024-03-29
MMStar is a 1,500-question multiple-choice benchmark of vision-indispensable samples across six core capabilities, filtered so that questions cannot be answered from text alone or from training-data leakage. This row holds overall accuracy as reported by model developers under their own harnesses; it is a component of the fixed-harness OpenVLM family (mmstar-vlmevalkit). Higher is better.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen3.8-Max | 85.9 |
| Gemini 3.1 Pro Preview | 84.0 |
| Qwen3.5 397B A17B | 83.8 |
| Qwen3.6 Plus (2026-04-02) | 83.3 |
| Qwen3.7-Plus (source unspecified) | 83.2 |
| Gemini 3 Pro Preview | 83.1 |
| Qwen3.5 122B A10B | 82.9 |
| GPT-5.6 Sol | 82.5 |
| Qwen3.5 35B A3B | 81.9 |
| Qwen3.6 27B | 81.4 |
| Qwen3.5-27B | 81.0 |
| Qwen3.6 35B A3B | 80.7 |
| Kimi K2.5 | 80.5 |
| Qwen3.5 Plus (2026-02-15) | 80.3 |
| Qwen3.5-9B | 79.7 |
| Qwen3.5-Omni-Plus | 79.4 |
| Qwen3 VL 32B Thinking | 79.4 |
| Qwen3 VL 235B A22B Thinking | 78.7 |
| Qwen3-VL-235B-A22B-Instruct | 78.4 |
| Qwen3.5-4B | 78.3 |
| InternVL3.5-241B-A28B | 77.9 |
| Gemma 4 12B IT | 77.7 |
| Qwen3 VL 32B Instruct | 77.7 |
| Gemini 2.5 Pro | 77.5 |
| Gemma 4 31B IT | 77.3 |
| GPT-5.2 | 77.1 |
| Opus 4.8 | 76.7 |
| GPT-5 | 76.4 |
| Gemini 2.5 Flash | 75.8 |
| Qwen3.5-Omni-Flash | 75.7 |
| Qwen3 VL 30B A3B Thinking | 75.5 |
| GLM-4.5V | 75.3 |
| Qwen3 VL 8B Thinking | 75.3 |
| Qwen3 Omni 30B A3B Thinking | 74.9 |
| LFM3-VL-2B | 74.4 |
| GPT-5 Mini | 74.1 |
| Sonnet 4.5 | 73.8 |
| Opus 4.5 | 73.2 |
| Qwen3 VL 4B Thinking | 73.2 |
| Gemini 3.5 Flash-Lite | 73.1 |