← Back to live feed · 1 stories across 1 day
Sunday, Sep 20, 2026
1 story1 Alibaba Ships Qwen Image 2.1 7B Open Weight Model for Unified Image Generation and Editing AI Sep 20, 9:06 AM EDT 23/13
A single checkpoint from Alibaba now handles both the creation and modification of imagery in a new open weight release. This tool, called Qwen-Image-2.1, uses a lightweight 7B architecture that natively generates and edits RGBA layers to support transparent images. It supports high-fidelity editing using up to 10 reference images and improves visual quality for portraits and text rendering. The software is distributed under a strict non-commercial license.
The model received immediate support from vLLM, ComfyUI, and Diffusers. In SGLang-Diffusion benchmarks, a single RTX 4090 24GB GPU generated a 1024x1024 image in 18.7 seconds and performed image editing in 21.7 seconds. This implementation uses native precision with CPU offload and requires 22.7 GiB of peak GPU memory during requests.