Qwen-Image-2.1 has been open-sourced: its 7B model supports image generation, editing, and transparent image output.

Beating AI News: Qwen releases Qwen-Image-2.1 and makes its model weights available. The new model integrates text-to-image and image editing capabilities into a single architecture, with its visual generation component boasting only 7 billion parameters, while natively supporting 2K-resolution images and transparent-background assets. Editing capabilities are the key highlight of this update. Qwen-Image-2.1 can accept up to 10 reference images at once, enabling users to recombine multiple characters or products into a single image. For local edits, users can directly select areas, paint, or provide masks, eliminating the need to run a separate editing model. It can also directly generate RGBA images with alpha channels, meaning the model outputs transparent-background assets natively, and can extract subjects from photos. Diffusers, ComfyUI, vLLM-Omni, and SGLang were all integrated on launch day, so local workflows don’t need to wait for slow third-party adaptations. However, the "open-source" tag here requires a caveat: while the weights are publicly available, they are licensed under the Qwen Research License, which permits non-commercial research and evaluation only; commercial use requires separate authorization from Qwen.