Qwen-Image 2.1
Qwen-Image 2.1 is a downloadable image model that generates and edits pictures, including ones with a transparent background, in a single 7B model.
New pictures and edits are the same model
Qwen released the weights on September 20, 2026. The visual part is 7 billion parameters, 32 layers. A separate Qwen3-VL 8B model reads the instruction and any images you attach.
- Text to image, or an edit of an existing image, through one pipeline.
- Transparent output is native. The README's suggested prompt says the image has an alpha channel and the background is transparent.
- Up to 10 reference images in one edit. Local changes can be marked with a circle, a painted note, or a separate mask.
- Default size is 2048 by 2048, 40 denoising steps. Other ratios go up to about 2752 on the long side.
Where people actually run it
- Diffusers, as QwenImage21Pipeline, on the day of release.
- ComfyUI, with separate text-to-image and editing workflow files.
- vLLM-Omni, SGLang, and LightX2V also shipped support the same day.
- Optional prompt-rewrite checkpoints, Qwen-Image-2.1-PE-T2I and PE-I2I, turn a short prompt into a longer one before generation. They are separate downloads.
Non-commercial use only
The file is the Qwen Research License, dated September 20, 2026, from Hangzhou Tongyi Laboratory Technology Co., Ltd. It defines non-commercial as research or evaluation only, and says commercial use needs a separate license from them.