Qwen releases Qwen-Image-2.1 weights: a unified model for image...

Qwen releases Qwen-Image-2.1 weights: a unified model for image generation and editing

The Qwen team (Alibaba) has released the weights for Qwen-Image-2.1 — a unified 7-billion-parameter model for image generation and editing with native RGBA transparency support and multi-reference editing, published on Hugging Face and ModelScope under the restrictive Qwen Research License.

Compiled by Sergey KostenchukPublished 2026-09-20Updated 2026-09-20

2026-09-20 Multimodal HuggingFace

🎨 Qwen releases Qwen-Image-2.1 weights: a unified model for image generation and editing

Qwen has made the weights for Qwen-Image-2.1 publicly available — a 7-billion-parameter model that natively works with RGBA transparency: it creates images with an alpha channel, edits transparent layers, and cuts out objects from RGB photos.

🌍 A single checkpoint combines generation, RGBA transparency, and editing that preserves faces and product textures — previously, transparency required a separate model. KV-cache reduces the cost of multi-image editing, but the Qwen Research License will hold back commercial adoption.

👤 The weights are available on Hugging Face; they can be run locally via diffusers (QwenImage21Pipeline, cpu offload for weaker GPUs) or in a demo Space. Up to 10 references and mask-based editing — for stickers, virtual try-on, and infographics. For commercial use, check the license.

Source 1: https://qwen.ai/blog?id=qwen-image-2.1 Source 2: https://huggingface.co/Qwen/Qwen-Image-2.1

Sources