Qwen2 VL 2B Instruct
Alibaba · 2024-08-29 · 2.2B parameters
Qwen2-VL-2B-Instruct is the smallest instruction-tuned vision-language model of Alibaba's Qwen2-VL series, released on August 29, 2024 with 2,208,985,600 parameters. It pairs a Qwen2-1.5B-class language backbone with the Qwen2-VL vision encoder, using naive dynamic resolution and multimodal RoPE (M-RoPE) for image and video understanding.
Benchmark scores
| Benchmark | Score |
|---|---|
| AI2D (test) | 74.7 |
| Capability | 104.8 |
| ChartQA (test) | 73.5 |
| DocVQA | 90.1 |
| EgoSchema | 54.9 |
| HallusionBench (average) | 41.7 |
| InfographicVQA | 65.5 |
| MATH-Vision | 12.4 |
| MathVista (mini) | 43.0 |
| MM-Vet | 49.5 |
| MMBench (Chinese, test) | 73.5 |
| MMBench (English, test) | 74.9 |
| MMBench v1.1 (English, test) | 72.2 |
| MME | 1872 |
| MMMU | 41.1 |
| MMMU Pro | 37.6 |
| MMStar | 48.0 |
| MVBench | 63.2 |
| Perception Test (MCQA) | 53.9 |
| RealWorldQA - release-reported visual perception | 62.9 |
| TextVQA (validation) | 79.7 |
| Video-MME | 55.6 |
| Video-MME (with subtitles) | 60.4 |