Alibaba's Qwen releases a 7B image model with open weights

Alibaba's Qwen team released Qwen-Image-2.1 on Sunday with open weights, describing it as a unified model for both generation and editing in a lightweight 7B package. The company says it natively generates and edits RGBA layers, supports up to 10 reference images with precise local control, and improves text rendering and portraits over its predecessor.

Third parties confirmed the release within the hour. ComfyUI said the model is supported, describing one 7B checkpoint that generates and edits, with image generation at native 2K and instruction editing from up to 10 reference images in a single pass. The vLLM project said it has day-0 support in vLLM-Omni, and put the architecture at a 7.1B diffusion transformer paired with Qwen3-VL-8B.

Qwen says the model outperforms most closed-source models and that inference on multi-image inputs is drastically accelerated. Those are the company's claims, and no independent benchmark has appeared. Local edits can be specified by circling regions, according to Qwen's examples, which show a watch removed, hair recoloured and clothing replaced in one pass.

The licence is the caveat. One developer who read it on Hugging Face describes it as strictly non-commercial, with no revenue cap. That would limit the model to research and personal use unless Alibaba grants separate terms, a point the launch thread does not address. The team had trailed the release ten hours earlier with a one-line post.