Qwen-Image 2.1: generation, editing and transparency in one model
Alibaba detailed the Pro and Turbo APIs for its image model: from $0.014 an image, up to 10 references and a real alpha channel.
Alibaba has published a detailed breakdown of Qwen-Image 2.1 and its APIs. The open weights came out on September 20, and the Pro and Turbo APIs went live in Model Studio on October 9. It’s one 7B-parameter model for generating images, editing them and cutting out subjects.
What’s new
- Transparency: the model generates images with a real alpha channel, edits transparent layers and extracts subjects from photos.
- References: up to 10 images in one pass, preserving the look of people and products.
- Local edits: it changes a marked region and leaves the rest of the image intact.
- Text: more legible headlines, labels and dense infographics.
- Resolution: native 2K, with 2048×2048 as the default.
Turbo runs 8 steps instead of 40 and suits high volume, while the company suggests Pro for hero images.
How to try it
In Model Studio, the models are qwen-image-2.1-turbo and qwen-image-2.1-pro. In Global regions, Turbo costs about $0.014 an image and Pro about $0.035. Singapore is higher at $0.016 and $0.04. Failed calls aren’t charged. For local evaluation, the weights are on Hugging Face and work with Diffusers and ComfyUI.
Limitations
The open weights carry a research license, for research and evaluation only. Commercial use requires Model Studio or a separate license from Qwen. The company itself calls its LMArena ranking preliminary, based on a September 22 snapshot.
Prepared with AI from public sources.