Alibaba releases Qwen3-VL with visual-agent and 3D grounding abilities

Background

Vision-language models accept images or video together with text and answer in text. Qwen3-VL was the September 2025 release in Alibaba's Qwen vision-language line. ^[2]

What happened

Alibaba released Qwen3-VL in September 2025 under Apache 2.0. The family includes dense and mixture-of-experts models in Instruct and Thinking editions, and its flagship is Qwen3-VL-235B-A22B. ^[2][1]

Key details

The model has a native 256K-token context, which can be expanded to 1 million tokens, and it can index hours-long video at second-level granularity. It acts as a visual agent that operates PC and mobile interfaces, and it can generate Draw.io diagrams and HTML, CSS and JavaScript from images or video. ^[1]

OCR coverage grew from 19 to 32 languages, with better handling of blurred, tilted and low-light text. ^[1]

Impact

The release shows that open multimodal models are moving from describing images toward acting on them, such as reading a screen, generating code from a visual design or locating events in video. Because the weights were released under Apache 2.0, these abilities can be tested and built on by anyone, not only by users of a hosted service. ^[1]

Several terms used to describe vision-language models are worth defining. Grounding is the task of connecting words to places in an image or video, such as finding the object a sentence describes. A visual agent is a model that reads what is on a screen and takes actions in an interface. OCR, optical character recognition, turns text in images into machine-readable text. Together these abilities make a vision-language model useful for documents, software interfaces and video, not only for captioning photographs.

The native context window is the amount of text, images or video the model can take in one pass. A longer window lets a model keep a whole lengthy document or video in view, rather than processing it in pieces and joining the results afterwards.