Qwen open-sources Qwen-Image-2.1, a 7B model that unifies image generation and editing with native transparency under a non-commercial license
- Qwen open-sourced Qwen-Image-2.1, an image model whose visual generation component has 7B parameters across 32 Single-Stream DiT layers, unifying text-to-image generation and image editing in one model.
- The release ships under a non-commercial license, with commercial use requiring a separate license from Qwen/Alibaba, while earlier Qwen image models from 2025 used Apache 2.0.
- The prompt decides whether the output is a regular image or one with an alpha channel, folding the transparency work of the December 2025 Qwen-Image-Layered model into the unified model.
- Editing accepts up to 10 reference images and can pull a subject out of an RGB photo as an RGBA layer, edit transparent layers, and rewrite text inside them (the example replaces BLOOM with Qwen-Image).
- Inference speed comes from mixed-granularity attention, with a token-level causal mask for text and a chunk-level mask for images, plus KV cache reuse that caches input images and editing instructions as static context in the first step.
Hacker News opinions
Local text-to-image is pretty damn impressive right now, and I'd say it's ahead of local code generation. I get an image in seconds at a quality way above what I expect from a local model, while coding is much slower and much less impressive.
I bet it's the reverse for an artist. They'd be less impressed by the image and more impressed by the code.
I've been running FLUX.1-Kontext on my M1 Max with mflux, about 3 minutes per 512x512 edit. It takes a photo of a person or an animal and gives back a line drawing that's recognisably them instead of a generic similar face, and setup is far simpler than comfyui.
Image gen you eyeball one frame and stop. Code needs hundreds of tokens all correct in sequence, one bad line and the whole thing fails.
In the Cheers team example the otherwise great result genericizes Shelley Long.
The source image wasn't great so it's a decent representation. Woody Harrelson comes off much worse.
A 7B model with these capabilities is impressive and kind of worrying. Does Qwen do any watermarking yet?
They've had a mark in Fourier space on their models for a while, and even without it VAEs are pretty easy to detect.
The Qwen team ships the most diverse set of models of any Chinese lab. Only Gemini/DeepMind comes close.
How do I run this locally the way I run llama-server -m model, without Python in the loop?
ComfyUI is the easiest way into local image or video models, or vLLM if support is already there: vllm serve model --omni --port 9080.
The GitHub page lists Diffusers, ComfyUI, vLLM-Omni, SGLang and LightX2V with links for each.
Look at diffusion.cpp, it's built for these models. I got krea-2-turbo running with it two months ago.
A 7B diffusion model now renders CJK text better than Microsoft Windows.
That feature only landed in the official ChatGPT image gen recently, and now it runs locally. Blows my mind.
I'm grateful to the Chinese labs for open sourcing their best models. If it was left to the Americans we'd be paying obscene API fees.
Check the license though: no commercial use of the Materials without a separate commercial license from them. Cheaper than other image models maybe, but that's up to Qwen/Alibaba's whims, not cloud pricing.
Earlier Qwen models were Apache licensed, this one looks much more restrictive.
The last image models with Apache 2.0 were from 2025. Recent Qwen models are non-commercial use.
I'll bet a nonzero amount of its training data is GPL, so I'll treat it as GPL and use it however I want. If AI labs ignore licenses, so do we.
They finally fixed their VAE. It held their models back for two years.
Happy to see a new open image model from Qwen, but the license is a letdown, and it doesn't even beat their closed Qwen image model, which is already a bit old.
Qwen Image 3 came out two months ago. I think you're confusing it with another model.