What it does
Qwen3-VL is a series of vision-language models developed by the Qwen team at Alibaba Cloud and published under the Apache License 2.0. The repository calls it the most capable vision-language line in the Qwen series so far, with gains in text understanding, visual reasoning, context length, video and spatial understanding, and agent use. The models come in dense and mixture-of-experts designs that scale from edge to cloud. Each size has an Instruct edition and a reasoning-focused Thinking edition, and the release notes list sizes from 2B up to 235B-A22B, plus FP8 versions. The README lists a native 256K context expandable to 1M, OCR in 32 languages, a visual agent that operates PC and mobile interfaces, and generation of Draw.io, HTML, CSS and JavaScript from images or video. It also covers 2D and 3D grounding for spatial tasks. Usage examples run through Hugging Face Transformers and ModelScope, with cookbooks for tasks such as document parsing, OCR, video understanding and computer-use agents.
Capabilities
- Dense and mixture-of-experts sizes from 2B to 235B-A22B
- Instruct and Thinking editions of each model
- Native 256K context, expandable to 1M
- OCR across 32 languages and long-document parsing
- Visual agent for operating PC and mobile interfaces
- Code generation from images and video
When to use it
- Describing, comparing or answering questions about images and video
- Extracting text and structure from scanned or photographed documents
- Building agents that read screens and control software interfaces
- Running an open vision-language model on your own infrastructure
