Qwen releases Qwen-Image-2.1, a 7B open-weight model that generates and edits transparent images from up to ten references

Alibaba's Qwen team released Qwen-Image-2.1 on September 20, an open-weight model that combines text-to-image generation and image editing in a single workflow. Its visual generation component has 7 billion parameters arranged as 32 single-stream diffusion transformer layers. Qwen says a lightweight architecture using mixed-granularity attention and prefix KV cache reuse keeps computational cost low, particularly when several reference images are supplied, and The Decoder reports it runs on capable consumer GPUs such as an RTX 3090. The release natively generates and edits transparent RGBA images, so users can isolate objects, edit transparent layers, or change text on them. It accepts up to ten reference images at once for uses such as group portraits, virtual try-ons and room design, and local edits can be guided by circles, painted annotations or separate masks while preserving the identity of people and products. Qwen also claims improved typography, portrait lighting and fine detail, with native 2K output across multiple aspect ratios. Weights are available on Hugging Face, ModelScope and GitHub, with a Hugging Face demo. Diffusers supports the model from day zero through a dedicated pipeline, and ComfyUI ships compatible weights and example workflows. Qwen says the model beats most closed models on its own benchmark, but independent results are still pending. The research license bars commercial use, so businesses must apply to Qwen for a separate license. The release follows the Qwen3.8 Omni-Flash and LiveTranslate models covered in the previous edition.