On September 20, 2026, Qwen (Tongyi Qianwen), Alibaba's AI development team, released Qwen-Image-2.1, an image generation AI. It combines the ability to create images from text and the ability to edit images in a single model, and the model weights (the trained data) are distributed on Hugging Face and ModelScope. It is a so-called "open-weight" model that you can download and run in your own environment with a GPU.
The announcement
Qwen-Image-2.1 is notable for keeping the image-generating part relatively small, at 7 billion parameters (7B). The Qwen team describes it as a model that balances generation quality, processing efficiency and versatility.
At release, Diffusers, an image generation library, ComfyUI, a tool for chaining image generation tasks together as nodes, and the inference software vLLM-Omni, SGLang and LightX2V for running it fast all supported it from day one. It was released with the surrounding tools in place so it can be built into local environments and servers right away.
The Qwen team lists the following four main improvements in this release.
- A light, efficient design that delivers high image quality with less computation
- Generating images with transparent backgrounds and regular generation and editing in one model
- Wide-ranging editing using up to 10 reference images
- Better rendering of textures and text, and of lighting in portrait photos
Transparent image generation and wide-ranging editing
The most eye-catching feature is that it can create images with transparent backgrounds (RGBA images, which carry transparency information in addition to color information) directly from text instructions. Until now, a separate step was needed to cut out the background after creating an image, but Qwen-Image-2.1 can output a transparent background from the start. This can reduce the work of creating stickers, icons and product image assets.
The team says the same model can also edit elements within transparent images and cut out only people or objects from photos. ComfyUI has provided templates (workflows) for generating images from text and for editing images.
For editing, up to 10 reference images can be used at the same time. The Qwen team shows an example of creating a group photo from the face photos of six people. It says edits can keep the appearance of people and products intact.
When you want to fix only part of an image, you can specify the area by circling it, painting over it with color, or passing a separately prepared mask (an image that shows the area to edit).
Resolution and performance
The output resolution is 2048×2048 pixels for square (1:1) images. It supports seven aspect ratios from portrait to landscape, and at 16:9 it can output 2752×1536 pixels.
On performance, according to DataNorth, an AI consulting firm in the Netherlands, Qwen-Image-2.1 scored 60.28 on Qwen-Image-Bench, an evaluation framework Qwen built itself. It slightly exceeded Google's Nano Banana 2.0 (59.82) but fell short of Qwen Image 3 Pro (62.36), a higher-end Qwen model whose weights are not released, and trailed OpenAI's GPT Image 2.5 Sunburst (67.01), which ranked first. All of these are evaluations by Qwen itself, not verified by a third party. In addition, guidance such as how long generation takes and how much GPU memory is needed has not been officially provided.
How to run it and the license
The basic approach is to load it from Diffusers, and the sample code Qwen provides sets the image-creating process to 40 steps. A method for moving part of the processing to the CPU when GPU memory is insufficient is also described. A demo for trying it out is also available on Hugging Face.
The license deserves attention. Qwen-Image-2.1 is distributed under the Qwen Research License, and use is limited to research and evaluation purposes. Commercial use requires applying to Qwen and obtaining a separate commercial license. The original Qwen-Image was released under the Apache 2.0 license, which also allows commercial use, so people who have been using Qwen's image models for work need to note that the terms have changed.
Subsequently, on October 9, the Qwen team released the weights for Qwen-Image-2.1-Turbo, a fast version that can create images in 8 steps, and the Pro and Turbo versions of Qwen-Image-2.1 can now be called via API on Alibaba Cloud's Model Studio.
CoAI's take
What is new about Qwen-Image-2.1 is that it brings direct creation of images with transparent backgrounds and editing with multiple reference images together in a 7B model you can run locally. It could be handy for work that involves a lot of cutting out and compositing, such as creating design assets and product images for e-commerce sites. On the other hand, because the license has changed to research use, you are free to try it, but you cannot use it as is for production work. If you want to use it commercially, you will need to consider an individual agreement with Qwen or using the Model Studio API. Chinese AI companies continue to release open-weight models, and this is an example showing that license terms need to be checked model by model.
