One smaller model for making and editing images
Qwen-Image-2.1 handles both text-to-image generation and image editing in one model, with 7 billion parameters in its visual generation component across 32 single-stream DiT layers. It generates and edits transparent images natively, folding in a capability Qwen had shipped as a separate model, Qwen-Image-Layered, in December 2025. It accepts up to 10 reference images, and edits can be aimed with circles, painted annotations or a separate mask. Qwen says a mixed-granularity attention design and KV cache reuse make multi-image editing faster. [1]
The 7 billion figure covers the diffusion transformer only; the pipeline also loads Qwen3-VL 8B as its text encoder. MarkTechPost notes that the original Qwen-Image, released in August 2025, was a 20 billion parameter model under Apache 2.0, with editing in a separate checkpoint. On Qwen-Image-Bench, Qwen's own benchmark, the new model scores 60.28, above every open-weight model on Qwen's chart, with six closed models scoring higher. [3][2]
Open weights, closed to business
The weights ship under the Qwen Research License Agreement, dated September 20, 2026. It grants use for non-commercial purposes only, which it defines as research or evaluation, and says commercial use requires a separate license requested from Qwen. [4]
Atlas interpretation: Qwen's announcement calls the release open-source, and anyone can download the weights, but the license is the fact that matters to anyone building a product: this model cannot be deployed commercially without a separate deal. For image generation, Qwen's openness has narrowed since the original Qwen-Image, released under Apache 2.0 in August 2025. [1][4][2]
Sources
- Qwen-Image-2.1: Compact, Efficient, and Unified Image Creation
Qwen · Sep 20, 2026
- Alibaba Qwen Releases Qwen-Image-2.1: A 7B Open-Weight Model for Image Generation and Editing
MarkTechPost · Sep 21, 2026
- QwenLM/Qwen-Image-2.1
GitHub · Sep 22, 2026
- Qwen RESEARCH LICENSE AGREEMENT
GitHub · Sep 20, 2026