Alibaba has released Qwen Image 2.1, a new image generation model featuring 7 billion parameters. The model supports native transparency and allows users to composite images using up to 10 reference images to maintain consistency.
In internal benchmarks, the model scored 60.2, trailing OpenAI GPT Image 2.5 Sunburst at 67 and Muse Image at 62.34, but outperforming Google Nano Banana 2.0, which scored 59.82. On the AI Arena platform, it achieved a score of 1228.
The model is designed to run on consumer hardware. Early adopters report generating 1MP images in approximately five seconds on an RTX 4090 and about 25 seconds on RTX 5070 or 5080 cards. Users also report running it on an RTX 3060 with 64GB of memory.
The licensing terms for Qwen Image 2.1 restrict commercial resale of the model itself, requiring a separate license from Alibaba for such use. However, the developer clarified that users retain the rights to the images generated by the model.
