September 20, 2026
Qwen-Image-2.1 weights are open, but not for profit: the license permits research only
Alibaba released Qwen-Image-2.1 7B weights on 20.09, but commercial use is closed: research and evaluation only.
Qwen
@alibaba_qwen
Meet Qwen-Image-2.1, the most balanced and cost-effective image-generation model in the Qwen-Image series! Now with open weights! 🎨 One model for both generation and editing: top-tier quality in a lightweight package. Highlights: 👀 - Compact and very fast: its lightweight 7B architecture outperforms most closed models, while multi-image inference has been dramatically accelerated. - Native transparency: generates and edits RGBA layers, making compositing and text editing within transparent images seamless. - Flexible, high-precision editing: up to 10 reference images and precise local control preserve the likeness of portraits and products. - Broad coverage and beautiful visuals: panoramas, infographics, virtual try-ons, realistic textures, and elegant typography. Create your next masterpiece with Qwen-Image-2.1! 🖼️ - Blog: https://qwen.ai/blog?id=qwen-image-2.1 - GitHub: https://github.com/QwenLM/Qwen-Image-2.1 - ModelScope: https://www.modelscope.cn/models/Qwen/Qwen-Image-2.1 - Hugging Face: https://huggingface.co/Qwen/Qwen-Image-2.1
· 2.1M views
Open Qwen-Image-2.1 weights can be installed with four commands. They cannot be used commercially.
The model is licensed not under Apache 2.0 but under the Qwen Research License. Section 2.a permits use only for non-commercial purposes; Section 2.b directs users seeking a separate commercial license to model-business@notice.qwencloud.com.
It used to be different. The first Qwen-Image launched on 04.08.2025 with 20B parameters under Apache 2.0, and that license allowed paid products to be built on it. Version 2.1 is three times lighter on the visual side, while the door to commercial use is now separate and handled by email.
You can try the model without your own GPU in the official Hugging Face Space: Qwen/Qwen-Image-2.1; at the time of checking on 20.09, it was running.
Install it locally. Four pip commands, then the QwenImage21Pipeline class. The model card includes ready-made snippets for generation, editing, and transparent RGBA images.
The environment is unusual: diffusers is needed from a git branch, transformers from 5.17, and torch from 2.4.0. If video memory is short, pipe.enable_model_cpu_offload() helps.
Editing accepts up to 10 reference images at once. Local edits are specified with a circle over the image, an annotation, or a mask, and the model keeps people’s faces and product appearance unchanged.
Transparency is native in 2.1. The model outputs RGBA layers itself; the exact prompt format is specified in the Hugging Face card.
Support landed in diffusers on 18.09.2026 (PR #14804); there is no package release with the model yet—it is installed from a git branch.
