Alibaba released Qwen-Image-2.1 on September 20, 2026, saying it can generate images that outperform most closed models on its own benchmark. It is the company's first major update to its image generation model since its initial release.
Alibaba reported the model has 7 billion parameters, measured on its internal benchmark, the Qwen benchmark. That compares with the previous version, which had a higher parameter count.
Qwen-Image-2.1 is built on an optimized architecture and targets image generation and editing. Availability begins on Hugging Face, GitHub, and Model Scope, initially for researchers and developers.
"This group portrait was assembled from individual photos of each person," said Matthias Bastian, a contributor at thedecoder. The model natively generates and edits transparent images (RGBA), letting users isolate objects or change text on transparent layers.
The announcement follows Alibaba's broader push into open-source AI development. Alibaba framed the release as a step toward democratizing access to advanced image generation technology.
Alibaba did not say how the model performs on external benchmarks, and it raised questions about the limitations of its internal testing. The company said it is working on additional benchmarks and improvements.
Source: thedecoder
