Qwen open-sources Qwen-Image-2.1

Alibaba's Qwen team announced Qwen-Image-2.1, an open-source image model that puts text-to-image generation and image editing in a single system. It follows Qwen-Image-2.0 and the earlier Qwen-Image-Edit-2511 editor.

The model's visual generation component has 32 single-stream DiT layers and 7B parameters. Qwen says the design aims to balance image quality against compute cost. It published a comparison on its own Qwen-Image-Bench against open- and closed-source models, but the blog text does not give the scores.

Most of the efficiency gains are for editing with several input images. Text, including the system prefix and editing instructions, uses a token-level causal mask, while image generation uses a chunk-level mask. Input images and instructions are computed once in the first step and kept in a KV cache, which Qwen says speeds up inference and reduces memory use.

The main new feature is native transparency. Qwen-Image-Layered, released in December 2025 as a separate model, generated transparent images. Version 2.1 builds that capability into the main model, which uses the prompt to decide whether to output a regular image or one with an alpha channel. It can also edit transparent layers, including the text inside them, and extract a subject from a regular RGB photo as an RGBA layer.

Other editing changes:

• Multiple references: up to 10 reference images can be combined, for example for group portraits, virtual try-on or room layouts.
• Local edits: the area to change can be marked with circles, painted annotations or a separate mask.
• Fidelity: faces and product details such as text, texture and shape are kept more consistent across edits.
• New tasks: panoramas, infographics and storyboards.

The announcement does not state the licence, where the weights are hosted, or any independent benchmark results.