Qwen open-sourced Qwen-Image-2.1's weights, but the license is research-only where Qwen-Image was Apache 2.0
The 7B model makes transparent images straight from a prompt, and it takes up to 10 reference pictures for a single edit.
Qwen put the weights for Qwen-Image-2.1 on Hugging Face on September 20, 2026, about ten hours after its developer account said they were coming. It's a 7B model in BF16, and the card calls it one model for making images and editing them.
Scroll to the bottom, though, and there's a license line. Qwen-Image-2.1 is under the Qwen Research License Agreement, and Qwen-Image (the 20B model on the same Hugging Face account) is under Apache 2.0.
- What
- Unified text-to-image generation and image editing model
- Size
- 7B parameters in its visual generation component
- License
- Qwen Research License Agreement
- Editing
- Support up to 10 reference images
What the card describes
Generation runs 7B parameters across 32 Single-Stream DiT layers, and Qwen lists four improvements. Transparency is the one that surprised me. Ask for an RGBA image in the prompt and you get a sticker back with nothing behind it, out of the same model that does ordinary text-to-image.
Editing takes up to 10 reference images. You mark what you want changed by circling it or painting over it, or you hand the model a separate mask, and Qwen says faces and products keep their identity, though the card doesn't say how that was measured.
Efficiency is the other claim. Qwen names mixed-granularity attention and prefix KV cache reuse (the card explains neither), so the whole memory section comes down to a single line that offloads the pipeline to the CPU.
We are excited to open-source Qwen-Image-2.1
Typography and portrait lighting are harder to check off a page. Qwen's examples name a group photograph built from six portrait references, plus a text rendering sample. Nowhere does Qwen print a benchmark score or a date. You do get code: two worked examples, and seven aspect ratios that top out at 2752 by 1536.
Three days from the tease
On September 17, 2026, the Qwen Developers account offered 50 people a look at the model before anybody else got one.
Qwen-Image 2.1 is going open source and we’re opening up 50 early access spots for you to try it out before release!
Late on September 19, Eastern time, the same account put a clock on it.
Qwen-Image-2.1 open source in 10hrs
The license line
Each card names its license in a single sentence and stops there. The older page says "Qwen-Image is licensed under Apache 2.0." The new one says the model "is licensed under the Qwen Research License Agreement".
On September 20, 2026, the Qwen-Image-2.1 page's "Downloads last month" counter read 183.