Alibaba's Qwen team open-sourced the image model Qwen-Image-2.1 on September 20, folding text-to-image generation and image editing into a single checkpoint. Weights are up on Hugging Face, with day-one support in diffusers and ComfyUI, plus a browser-based demo that needs no installation.
The biggest change from earlier generations sits in the license. The README specifies the Qwen Research License Agreement, which permits research and evaluation use only; deploying it in a commercial product requires applying separately to the Qwen team for a commercial license.
A 7B Model for Both Generation and Editing
Per the official README, the visual generation module is a 7B-parameter, 32-layer single-stream DiT. The text encoder is Qwen3-VL 8B, which encodes text instructions and reference images into the same representation space. The autoencoder is 64-channel RGBA with 16x spatial compression, so images with transparent backgrounds can be generated and edited natively, without first rendering a white background and cutting it out.
Native resolution goes up to 2K across seven recommended aspect ratios: 2048×2048 for 1:1, 2752×1536 for 16:9, and 1536×2752 for the portrait 9:16. On the editing side, a single pass can take up to 10 reference images, and localized edits can be marked with circles, scribbles, or a separate mask; the appearance of people and products is meant to stay consistent across multiple editing rounds. Official demos also cover panoramas, infographics, storyboards, and pulling a subject directly out of a photo.
Two design choices drive the speed. One is mixed-granularity attention: text uses a causal mask at the token level, images use a bidirectional mask at the patch level. The other is prefix KV-cache reuse — the input image and instructions are computed once at the first denoising step, then the cache is reused for the following dozens of steps. Alibaba recommends 40 inference steps and hasn't published specific generation times or minimum VRAM, only noting that cards without enough memory can turn on CPU offload.
Benchmark Scores and Standing
Third-party Qwen-Image-Bench data puts Qwen-Image-2.1 at a total score of 60.28, edging out Nano Banana 2.0's 59.82 and GPT Image 1.5's 59.65; the previous-generation Qwen-Image 1.0 scored 49.23. On the full leaderboard of 29 models it ranks seventh — the top six are all closed-source, with GPT Image 2.5 Sunburst leading at 67.01. This is Alibaba's own evaluation set, and independent reproduction hasn't surfaced yet.
Day-one inference framework support is broad: vLLM-Omni, SGLang, and LightX2V all list Day-0 compatibility. Judging by the model card, Alibaba wants this integrated into developer workflows as quickly as possible.
From Apache 2.0 to a Research License
Tracing the licensing history of the Qwen-Image line makes the shift clear. Qwen-Image 1.0 and Qwen-Image-Edit from August 2025, and Qwen-Image-Layered and Qwen-Image-2512 from December, all shipped under Apache 2.0, with no strings attached for commercial use. With 2.1, the license switches to a research license; the text states that the materials may not be used for any commercial purpose without a separately obtained commercial license.
For smaller teams doing e-commerce photo retouching or poster generation, that's a real barrier. The previous generation could be deployed straight into a paid product; this one can only be used for internal testing until a license is secured. The license doesn't spell out any revenue or user-count exemption, and Alibaba hasn't published pricing or made any public statement about how licensing fees will work.
Qwen's large language model side has already made a similar move — this site previously reported on Qwen3.6-Max going closed-source. The image model still ships open weights this time, but it's closed off commercial use halfway. Whether the community workflows, LoRAs, and plugins built around the Apache-licensed versions can carry over to commercial use on 2.1 will depend on Alibaba filling in the licensing terms.
Sources: Qwen official blog, Qwen-Image-2.1 GitHub README and Hugging Face model card (parameters, resolution, license), ComfyUI official blog, CocoLoop, CellCog license comparison roundup (licensing history across Qwen-Image generations and Qwen-Image-Bench scores).