Qwen3-VL

Background

Qwen3-VL was released in September 2025 by Alibaba's Qwen team as a vision-language model, meaning a model that takes both images or video and text as input and produces text. It was released under Apache 2.0, a permissive license. ^[2][1]

What it is

The family includes dense and mixture-of-experts models, each offered in an Instruct edition and a reasoning-enhanced Thinking edition. The flagship is Qwen3-VL-235B-A22B, which has 235 billion total parameters, with about 22 billion active per token. ^[1]

Key abilities

The model has a native context of 256K tokens, expandable to 1 million. Within that window it can handle hours-long video, indexing events at second-level granularity. ^[1]

It can act as a visual agent that operates PC and mobile interfaces. It can also generate Draw.io diagrams and HTML, CSS and JavaScript code from images or video. A visual agent is a system that reads what is on a screen and takes actions in response, such as clicking or typing. ^[1]

Qwen3-VL adds 2D and 3D grounding, which means locating objects in an image and describing their position in space. It can also judge position, viewpoint and occlusion, meaning whether one object hides another. ^[1]

Its OCR, the recognition of text in images, grew from 19 to 32 languages, and it handles text that is blurred, tilted or photographed in low light more reliably. ^[1]

Architecture

The model uses interleaved MRoPE positional embeddings, which are a way of telling the model where each token sits in a sequence of text, images and video frames. It applies DeepStack fusion, combining features from several levels of the vision transformer (ViT) rather than only the last one. It also aligns text with timestamps so that events in a video can be localized in time. ^[1]

Benchmarks

On ScreenSpot-Pro, a benchmark of GUI grounding, which is locating elements on a screen, the 235B Instruct model scored 62, as reported by Alibaba. Benchmarks measure narrow abilities, so the score should be read as one indicator among many. ^[1]

Position in the field

Qwen3-VL shows the direction of open multimodal models: beyond answering questions about a picture, the model is designed to read interfaces, produce code from visual input and work with long video. Its open license means developers can run and adapt it, which is useful for building tools that need to understand screens and documents. ^[1]