The Decoder said Alibaba's Qwen team released Qwen-Image-2.1 on 20 September with a 7 billion-parameter visual generator, RGBA editing, up to ten reference images, and a research licence that bars commercial use. AI Weekly, timestamped the same afternoon, reported the Hugging Face and ModelScope push and the move off Apache 2.0.
The Decoder's Matthias Bastian reported Qwen-Image-2.1 on 20 September. The visual generation component has 7 billion parameters. The team claims that on its own benchmark the model beats most closed models; independent benchmarks are still pending.[1]
It natively generates and edits transparent RGBA images and takes up to ten reference images at once, The Decoder wrote. Weights are on Hugging Face, GitHub and ModelScope; a research licence bars commercial use, so businesses must apply to Qwen for a separate licence. AI Weekly, published 20 September, said the Qwen team pushed the same weights to Hugging Face and ModelScope and that licensing moved from Apache 2.0 on the earlier Qwen-Image line to a non-commercial Qwen Research License Agreement.[1], [2]
Neither The Decoder nor AI Weekly published an independent closed-model leaderboard score for Qwen-Image-2.1. The Decoder said those benchmarks are still pending. AI Weekly's alert credits huggingface.co via Hacker News for the shipping notice.[1], [2]
