Qwen (千问) open-sourced Qwen-Image-2.1, unifying generation and editing with a 7-billion-parameter visual-generation component and native transparent-image support.

The Qwen (千问) team at Alibaba (阿里巴巴) announced the open-source release of Qwen-Image-2.1 on September 20. The model combines text-to-image generation and image editing, with 7 billion parameters specifically in its visual-generation component, rather than a stated total parameter count for every component.

The release supports native generation and editing of transparent images alongside ordinary images. Its described functions include transparent-layer editing and cutout operations, rather than transparency being presented solely as a separate post-processing step.

The model supports up to 10 reference images, with the announcement describing improvements in localized editing and preservation of portrait and product details. It also claims improvements in text layout, lighting, and visual detail, while citing lightweight architecture and inference optimization as measures to balance quality and compute cost; no quantitative speed, memory, or cost benchmarks are supplied.

MetaX (沐曦股份) separately announced Day-0 adaptation for Qwen-Image-2.1 following the release. That establishes a reported compatibility milestone, but the supplied message does not name accelerator models, specify supported configurations, or provide performance measurements.