Qwen-Image-2.1 Ships With Native Transparency

Alibaba's Qwen team released Qwen-Image-2.1 on Sept. 20, an open-weight model that unifies text-to-image generation and editing with native support for transparent, RGBA output. The visual generation component holds 7 billion parameters across 32 single-stream diffusion transformer layers, according to the model's GitHub README.

The Qwen-Image-2.1 release lets a single model generate an image with a transparent background, then edit it, without switching to a separate tool for the transparency step. The model accepts up to 10 reference images for a single edit and supports local edits through circles, painted annotations or masks, the README said. It runs on a consumer GPU such as an RTX 3090, according to the-decoder, which covered the release the same day.

Native 2K output across seven aspect ratios

The model generates at up to 2048-by-2048 pixels and scales the same 2K budget across six other aspect ratios, mapping each to a fixed resolution rather than stretching a square output.

Aspect ratio	Resolution
1:1	2048 x 2048
4:3	2400 x 1792
3:2	2528 x 1696
16:9	2752 x 1536
9:16	1536 x 2752

Default inference runs 40 steps per image, and the repository recommends enable_model_cpu_offload() for GPUs without enough memory to hold the full model, the README said.

A research licence, not an open one

Qwen-Image-2.1 ships under the Qwen Research License Agreement, which bars commercial use, according to the license file linked from the GitHub repository. A business that wants to use the model in a product has to apply separately for a commercial grant, TechNode reported, placing Qwen-Image-2.1 alongside earlier Qwen releases that shipped weights but not unrestricted rights to them.

That licence sits at odds with how the release is being described. The GitHub repository calls the weights open, and Hugging Face and ModelScope both host them for direct download, but neither hosting page changes what the license file says a company may do with them.

The benchmark claim has no independent check yet

Qwen's team says the 7-billion-parameter model beats most closed competitors on Qwen's own benchmark. The-decoder reported that claim alongside an explicit caveat: independent benchmarks are still pending, and the comparison is Qwen's, run on Qwen's own evaluation set, not a third party's.

The GitHub README lists no benchmark table at all, comparative or otherwise, and names no specific closed model it claims to beat. That gap matters because DeepSeek's V4.1-Flash release topped its own charts before an outside lab found a gap between the vendor's numbers and what it measured independently, and China Telecom's Xing4.0 shipped with benchmark claims of its own that drew the same scrutiny.

Qwen-Image-2.1 arrives alongside Qwen3.8-Omni-Flash, part of what TechNode described as Alibaba's expanding open-source portfolio. Alibaba did not publish a technical report alongside the model, and the README does not name an individual researcher behind the release.

Five external projects added support on the same day the weights went up, the repository's news log shows: Hugging Face's Diffusers library, ComfyUI, vLLM-Omni, SGLang and LightX2V all shipped compatible pipelines on Sept. 20. That kind of coordinated day-zero support usually means Qwen briefed those projects before the public release, though neither the README nor Qwen's own blog said so directly.