Qwen2-VL
Qwen2-VL is Alibaba Qwen's second vision-language model family, reading images at any resolution and text in many languages.
A release of Qwen-VL. Other versions: Qwen3-VL, Qwen2.5-VL.
- Released
- August 29, 2024
- Model ID
- Qwen/Qwen2-VL-2B-Instruct, Qwen/Qwen2-VL-7B-Instruct, Qwen/Qwen2-VL-72B-Instruct
- Size
- 2B, 7B and 72B
- License
- Apache-2.0 for 2B and 7B. 72B was first offered only through an API
- Also searched as
- qwen2-vl, qwen2 vl, Qwen2-VL-7B-Instruct
What is new in Qwen2-VL
- Naive Dynamic Resolution: images of any resolution and aspect ratio map to a matching number of visual tokens.
- Multimodal Rotary Position Embedding (M-RoPE) for text, images and video.
- Reads text in images in many languages, including Japanese, Korean, Arabic and Vietnamese.
- Understands videos over 20 minutes long.
- Can act as a visual agent for phones and robots.
Qwen2-VL benchmarks
| Benchmark | Score | Note |
|---|---|---|
| DocVQA (test) | 94.5 | Qwen2-VL-7B. InternVL2-8B: 91.6 |
| TextVQA (val) | 84.3 | Qwen2-VL-7B |
| OCRBench | 845 | Qwen2-VL-7B. GPT-4o-mini: 785 |
| InfoVQA (test) | 76.5 | Qwen2-VL-7B |
Scores as published by the authors. We have not rerun them.
Use Qwen2-VL through an API
OCRRouter runs a different release of Qwen-VL as qwen/qwen-vl.
Qwen-VL versions
- Qwen-VL: all versions, price, speed and features on OCRRouter
- Qwen3-VL (September and October 2025)
- Qwen2.5-VL (January 28, 2025 (32B in March 2025))
- Qwen2-VL (August 29, 2024), this page
Questions
Sources
Last updated .