Qwen-Image-2.1 arrives with open weights
Qwen has released Qwen-Image-2.1, a 7B model with open weights that handles both image generation and editing in a single architecture.
The Qwen team has released Qwen-Image-2.1 with open weights. The model handles both image generation and editing within one architecture.
According to the announcement, it runs on a lightweight 7B architecture. The team says it outperforms most closed-source models and that inference is substantially faster when several images are fed in at once.
What the team lists
- Native transparency: generates and edits RGBA layers directly, allowing compositing and text editing inside transparent images.
- Up to 10 reference images: editing accepts as many as ten references with precise local control, and the team says fidelity is preserved for portraits and products.
- Coverage: described as strong on panoramas, infographics and virtual try-ons, with realistic textures and clean typography.
Qwen calls it the most balanced and cost-effective entry in the Qwen-Image series. Weights are available on Hugging Face and ModelScope, with code on GitHub.
No independent comparison of these claims has appeared yet; the figures and assessments come from the team that published the model.
See the LocalLLaMA thread and the Hugging Face page.
