How to run Alibaba's new image AI Qwen-Image-2.1 yourself

Alibaba's AI division Qwen released a new image model called Qwen-Image-2.1 on September 20, 2026. The model can both create new images from a text description and edit existing images, all in one and the same tool. It is free to download and can be obtained via Hugging Face, ModelScope or Qwen's own repository.

One of the new things about the model is that it can create images with a transparent background, as you know it from PNG files with transparency. It can also combine up to ten different reference images when you ask it to edit or compose a subject, and it can deliver images in high 2K resolution.

What is special about the news is that the model has only 7 billion parameters, which in the AI world is considered relatively small and easy to run. According to Alibaba's own tests, it still performs better than several closed, commercial models in the image space. Independent tests of the claim have not yet been published.

Because the model is open source, you can "self-host" it, meaning install and run it on your own computer instead of using a cloud service such as ChatGPT or Midjourney. This means you don't send your images or prompts to an external server, and that you are not dependent on an internet connection or subscriptions.

Several technology sites have already published step-by-step guides to the setup, including via the program ComfyUI, which provides a visual workspace, or via the Python library Diffusers for those who prefer code. For a good experience, a graphics card with at least 16 gigabytes of video memory (VRAM) is recommended, while weaker graphics cards can use a compressed version of the model.

It is worth noting that Qwen-Image-2.1 is released under a research license that does not readily allow commercial use. Companies that want to use the model in products or services must apply to Alibaba for separate permission.