Alibaba Releases Qwen-Image-2.1 in Non-Commercial Open Weights

Alibaba introduces Qwen-Image-2.1 in open weights for image generation and editing. The model features capabilities such as native RGBA and multi-reference, and it is touted as competitive against closed systems. However, its use remains limited to research without a dedicated commercial license, and independent evaluation is anticipated.

Available on Hugging Face, GitHub, and Model Scope, under research license

Qwen-Image-2.1 can be downloaded from Hugging Face, GitHub, and Model Scope, with a demo available on Hugging Face. Its distribution is under a research license that prohibits any commercial use. Interested companies must request a separate license from Qwen for professional use.

Native RGBA, ten references, and mask-guided editing

The model natively handles RGBA transparency, allowing for the isolation of objects or modification of text on transparent layers. It accepts up to ten reference images simultaneously for group portraits, virtual try-ons, or piece design, with local adjustments guided by circles, masks, or painted strokes. Its visual generation component boasts 7 billion parameters and can run on consumer graphics cards like the RTX 3090. Qwen reports inference gains related to architectural changes and KV cache reuse, especially in the presence of multiple references.

Claimed performance, third-party validation pending

Qwen claims that Qwen-Image-2.1 outperforms most closed models on its own benchmark. Independent evaluations are yet to come to confirm or refute these results.