Alibaba's Qwen team released Qwen-Image-2.1 on September 20, 2026: an image model that generates and edits images in one 7-billion-parameter generation component, with native support for transparent images. The weights are on Hugging Face, and the team describes the release as open source. The license says otherwise for business use: the Qwen Research License permits non-commercial use only, and commercial use requires a separate license from the licensor, Hangzhou Tongyi Laboratory Technology Co., Ltd.
The Qwen team's announcement lists four improvements:
On efficiency, the team describes a mixed-granularity attention design in which reference images and editing instructions are computed once and cached, reducing memory use and speeding up multi-image edits.
The Qwen-Image-2.1 model card states that the model is licensed under the Qwen Research License Agreement. That agreement grants rights to use, modify and redistribute the materials "for non-commercial purposes only," defines non-commercial as research or evaluation, and says commercial use requires requesting a separate license from the licensor, which the agreement names as Hangzhou Tongyi Laboratory Technology Co., Ltd.
So, despite the announcement's "open-source" wording, Qwen-Image-2.1 is best described as open-weight under a non-commercial research license. That matters for the most obvious use cases in the announcement, such as product imagery, virtual try-on, interior design and marketing assets, which are commercial. Teams can evaluate the model freely, but shipping its output in a commercial workflow needs a license conversation first.
The announcement compares Qwen-Image-2.1 with other open and closed image models on Qwen-Image-Bench, the team's own benchmark, and presents examples rather than third-party evaluations. A vendor benchmark and curated examples are a reasonable starting signal for an image model. They are not a substitute for testing on your own prompts, brand assets and editing tasks, and this article does not reproduce them.
Research teams and internal evaluators can use it under the license as published.
Design and e-commerce teams interested in transparent-asset generation or multi-reference editing should treat it as a candidate to evaluate, then contact the licensor named in the Qwen Research License about commercial terms before building on it.
Teams that need a permissive license today should check the license of every image model they consider, since terms vary widely across open-weight releases, as they do among the language models in this series, from the MIT-licensed DeepSeek-V4.1-Flash to the mixed licenses of Holo4.
When was Qwen-Image-2.1 released?
September 20, 2026, per the Qwen team's announcement.
Can I use Qwen-Image-2.1 commercially?
Not under the published license. The Qwen Research License Agreement allows non-commercial use only and directs commercial users to request a separate license.
What can Qwen-Image-2.1 do?
Text-to-image generation and image editing in one model, including transparent (RGBA) images and edits that combine up to 10 reference images.
Discover more content:
