Qwen-Image-2.1 Brings Native Transparent-Image Generation to a 7B Open Model
Qwen open-sourced Qwen-Image-2.1 on September 20, 2026, a compact image model that unifies text-to-image generation and image editing in a single 7-billion-parameter architecture, shrinking down from the 20B generator in its predecessor, Qwen-Image-2512.
Details
- Compact but competitive: built on 32 single-stream DiT layers with 7B parameters in the generator, scoring 60.28 on Qwen’s own Qwen-Image-Bench — ahead of every other model with downloadable weights, including Google’s Nano Banana 2.0 (59.82), though still behind Qwen’s own closed Qwen Image 3 Pro (62.36)
- Native transparency: generates RGBA images with alpha channels directly from text prompts, and can lift a subject out of an ordinary photo as a clean transparent cut-out — a task that previously needed a separate model
- Multi-reference editing: accepts up to 10 reference images in a single edit, supporting composites such as merging multiple portraits or product photos, guided by colored-circle selection, painted regions, or masks
- Efficiency architecture: uses mixed-granularity attention — token-level causal masking for text, chunk-level masking for images — plus KV cache reuse to cut redundant computation on repeated inputs and instructions
- Quality improvements: better typography rendering and more natural portrait lighting compared to the prior generation
- Licensing change: released under the Qwen Research License Agreement for non-commercial use only, requiring a separate commercial license from Alibaba — a shift from the Apache 2.0 terms used on several earlier Qwen releases
- Availability: weights published on Hugging Face, ModelScope, and GitHub as Qwen/Qwen-Image-2.1
What happened next
Qwen-Image-2.1 closes out a three-model release burst from Alibaba’s Qwen team in three days, following Qwen3.8-Omni-Flash and Qwen3.8-LiveTranslate. The more restrictive non-commercial license is a notable departure for a team that built much of its open-source reputation on permissive terms, and it will likely shape how quickly the model gets adopted outside research settings.