Qwen-Image-2.1

Qwen-Image-2.1 is a unified text-to-image and image-editing model from Alibaba's Qwen team, with weights published on 20 September 2026. Its visual generation component has 7 billion parameters across 32 single-stream DiT layers, and the vendor's own account of what is new is four things: a lighter architecture using mixed-granularity attention with the input images and instructions cached as static context, which matters most when editing from several references; native transparency, so the same model can output an ordinary image or an RGBA image with an alpha channel, edit a transparent layer, or cut a subject out of a photograph; editing from up to ten reference images, with the region to change marked by circles, painted annotations or a separate mask; and better typography and portrait lighting. Output runs to 2048 by 2048 at 1:1 and up to 2752 by 1536 at 16:9. The vendor publishes a comparison on Qwen-Image-Bench, its own benchmark, and only as a chart, so no score is recorded here. The fact a business needs before anything else is the licence, and the vendor's announcement does not mention it. Every earlier Qwen-Image model on the vendor's model-host account — six repositories, from the original in August 2025 to the December 2025 updates — was published under Apache 2.0. Qwen-Image-2.1 is published under the Qwen Research Licence Agreement, released the same day: use, modification and redistribution are granted for non-commercial purposes only, defined as research or evaluation, and commercial use requires a separate licence requested from the vendor by email. The licence is governed by Chinese law with the courts of Hangzhou as the exclusive venue, requires a "Built with Qwen" notice on any model trained with its outputs, and terminates for anyone who brings an intellectual-property claim against the licensor. The announcement nonetheless describes the release as open-source. A team already building on the Apache-licensed versions should treat 2.1 as a different legal object rather than a drop-in upgrade.