Qwen-Image-2.1 Released on HuggingFace: Image Generation and Editing with a 7B Model

The model unsloth/Qwen-Image-2.1 has been published on HuggingFace — a unified model for text-to-image generation and editing. Publication date — September 23, 2026. The model supports up to 10 reference images, local editing via masks and annotations, generation of transparent RGBA images, and extraction of objects from photos.

What it means

The visual part of the model contains 7.1 billion parameters, the architecture is a DiT with 32 Single-Stream layers. The largest weights file takes up 9.3 GB, with a total of 30.8 GB in the repository in safetensors format. It runs via diffusers. License — Qwen RESEARCH LICENSE AGREEMENT: non-commercial use only. The model may be used, modified, and distributed for research or evaluation. Commercial use requires a separate license from the rights holder. When distributing, you must include a copy of the license, retain the attribution notice, and indicate changes. If the model is used to train another AI model, the documentation must state "Built with Qwen" or "Improved using Qwen." The developer does not specify VRAM requirements on the model page. For local running, full weights in safetensors are available; there are no quantized builds in the repository. The model is aimed at research use and local experiments with image generation and editing. More news about AI tools for 3D and video — in the CPU3D news feed.