Benchmark
多模态安全榜
security
vision
multimodal
多语言
多模态安全榜：多维度评测综合成绩。评测分数为事实数据，整理自公开评测体系，仅供学习研究参考。
Top 10
榜单概览
数据说明
分数为 0-100 百分制综合成绩（Elo 榜为相对评分，仅供同榜内比较）。评测分数属事实数据，整理自公开评测体系与厂商公开资料，部分分数为厂商自报口径，不同榜单的评测方法与数据集可能存在差异，请结合多个榜单交叉参考。
Ranking
模型排名
| 名次 | 模型 | 机构 | 主分 | Toxicity | Legality | Privacy | Bias | Misinformation | Health Risk |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Sonnet 4 | Anthropic | 97.5 | 97.5 | 98.7 | 98.7 | 96.5 | 95.9 | 98.1 |
| 2 | GPT-4o | OpenAI | 93.2 | 94.4 | 95.9 | 92.7 | 87.5 | 93.1 | 95.4 |
| 3 | GPT-4.1 | OpenAI | 92.4 | 91.8 | 96.0 | 92.0 | 87.5 | 93.2 | 94.2 |
| 4 | GPT-4o mini | OpenAI | 91.6 | 92.0 | 94.4 | 91.1 | 83.7 | 93.9 | 94.4 |
| 5 | GPT-4.1 mini | OpenAI | 85.7 | 87.0 | 88.9 | 79.0 | 84.4 | 86.2 | 88.7 |
| 6 | Qwen2.5 VL 7B Instruct | Alibaba Cloud / Qwen Team | 81.8 | 80.1 | 79.4 | 77.1 | 86.5 | 84.7 | 83.1 |
| 7 | Gemma 3 27B | 79.6 | 77.8 | 79.4 | 63.7 | 90.1 | 83.7 | 82.7 | |
| 8 | GLM-4V-9B | Zhipu AI | 79.2 | 77.1 | 78.8 | 79.6 | 78.7 | 80.6 | 80.3 |
| 9 | Qwen-VL Max | Alibaba Cloud / Qwen Team | 79.0 | 75.9 | 74.6 | 71.5 | 82.6 | 88.8 | 80.8 |
| 10 | Mistral-Small-3.1-24B-Instruct-2503 | Mistral AI | 79.0 | 77.5 | 72.8 | 68.8 | 89.4 | 84.2 | 81.3 |
| 11 | Qwen2.5 VL 72B Instruct | Alibaba Cloud / Qwen Team | 78.8 | 69.9 | 73.1 | 69.2 | 93.9 | 83.8 | 83.0 |
| 12 | Gemma 3 12B | 78.6 | 82.2 | 78.1 | 62.9 | 88.1 | 78.4 | 82.1 | |
| 13 | Phi-3.5-vision-instruct | Microsoft | 78.2 | 90.5 | 84.5 | 90.5 | 57.5 | 58.7 | 87.4 |
| 14 | Gemini 2.5 Pro | 78.1 | 75.0 | 84.7 | 62.5 | 80.2 | 81.5 | 84.6 | |
| 15 | Gemma 3 4B | 77.7 | 76.4 | 76.7 | 64.3 | 85.1 | 82.1 | 81.7 | |
| 16 | Gemini 2.5 Flash | 71.7 | 68.0 | 68.2 | 61.4 | 74.8 | 80.9 | 77.1 | |
| 17 | Qwen2.5-VL-3B-Instruct | Alibaba Cloud / Qwen Team | 71.2 | 69.7 | 75.3 | 73.3 | 62.4 | 70.9 | 75.4 |
| 18 | InternVL3-78B-Instruct | Shanghai AI Lab | 69.2 | 69.3 | 66.0 | 66.9 | 69.5 | 75.5 | 68.1 |
| 19 | Qwen2.5 VL 32B Instruct | Alibaba Cloud / Qwen Team | 68.8 | 64.7 | 60.7 | 54.6 | 82.3 | 84.1 | 66.3 |
| 20 | LLaMA-3.2-11B-Vision-Instruct | Meta | 68.4 | 68.6 | 69.8 | 63.7 | 65.3 | 66.3 | 76.6 |
| 21 | InternVL3-14B-Instruct | Shanghai AI Lab | 65.2 | 68.1 | 67.2 | 65.0 | 58.9 | 66.3 | 65.8 |
| 22 | Grok-4 | xAI | 64.1 | 61.3 | 70.2 | 49.1 | 64.5 | 76.3 | 63.3 |
| 23 | QvQ-72B-Preview | Alibaba Cloud / Qwen Team | 63.5 | 69.6 | 56.8 | 61.4 | 67.4 | 60.0 | 65.9 |
| 24 | InternVL3-38B-Instruct | Shanghai AI Lab | 63.3 | 63.1 | 59.2 | 62.4 | 63.8 | 68.4 | 63.1 |
| 25 | LLaVA-Next-72B | LLaVA | 61.3 | 56.3 | 54.2 | 48.4 | 77.3 | 70.9 | 60.8 |
| 26 | Ovis2-34B | Alibaba Cloud / Qwen Team | 60.2 | 54.6 | 57.8 | 54.1 | 61.0 | 68.4 | 65.3 |
| 27 | InternVL3-8B-Instruct | Shanghai AI Lab | 60.0 | 65.3 | 58.4 | 54.1 | 61.0 | 62.8 | 58.6 |
| 28 | Ovis2-16B | Alibaba Cloud / Qwen Team | 58.2 | 53.9 | 52.7 | 54.1 | 60.3 | 68.9 | 59.5 |
| 29 | LLaVA-V1.6-Vicuna-13B | LLaVA | 56.3 | 54.9 | 49.3 | 40.1 | 69.5 | 66.8 | 57.4 |
| 30 | Ovis2-8B | Alibaba Cloud / Qwen Team | 55.5 | 55.6 | 48.2 | 48.4 | 66.0 | 57.7 | 56.9 |
| 31 | InternVL3-9B-Instruct | Shanghai AI Lab | 54.0 | 53.4 | 48.7 | 54.8 | 53.9 | 63.8 | 49.4 |
| 32 | MiniCPM-V-2.6 | OpenBMB | 52.1 | 56.5 | 45.2 | 48.4 | 56.0 | 58.2 | 48.5 |
| 33 | Pixtral-12B | Mistral AI | 50.9 | 48.5 | 45.1 | 42.0 | 61.0 | 55.6 | 52.9 |
| 34 | LLaVA-V1.6-Mistral-7B | LLaVA | 50.7 | 44.7 | 40.2 | 35.0 | 73.1 | 66.3 | 44.9 |
| 35 | Kimi-VL-A3B-Instruct | Moonshot AI | 50.5 | 56.5 | 47.1 | 42.7 | 51.8 | 57.7 | 47.1 |
| 36 | SAIL-VL-8B | ByteDance | 48.8 | 46.1 | 43.4 | 42.7 | 57.5 | 56.1 | 47.2 |
| 37 | LLaVA-Next-110B | LLaVA | 47.5 | 46.6 | 46.7 | 40.8 | 47.5 | 52.6 | 50.8 |
| 38 | InternVL3-2B-Instruct | Shanghai AI Lab | 47.5 | 44.2 | 38.9 | 42.0 | 58.2 | 56.1 | 45.4 |
| 39 | InternVL3-1B-Instruct | Shanghai AI Lab | 46.6 | 41.7 | 35.2 | 50.0 | 50.0 | 66.7 | 36.4 |
| 40 | DeepSeek VL2 | DeepSeek | 42.7 | 55.8 | 48.8 | 51.6 | 20.6 | 30.1 | 49.3 |
| 41 | GLM-4.1V-9B-Thinking | Zhipu AI | 42.4 | 40.4 | 36.6 | 30.6 | 54.6 | 53.1 | 39.2 |
| 42 | DeepSeek VL2 Small | DeepSeek | 37.4 | 43.7 | 44.7 | 47.8 | 18.4 | 24.0 | 45.5 |
| 43 | DeepSeek VL2 Tiny | DeepSeek | 35.6 | 40.4 | 35.0 | 37.6 | 27.0 | 32.7 | 40.8 |
| 44 | InstructBLIP-Vicuna-13B | Salesforce | 28.8 | 34.6 | 26.7 | 22.2 | 28.6 | 29.7 | 30.7 |
| 45 | InstructBLIP-Vicuna-7B | Salesforce | 28.6 | 29.3 | 25.0 | 22.3 | 31.2 | 31.1 | 32.9 |