Qwen-Image-2.1: a 7B open image model with a no-ship license
Native transparency, editing from 10 reference images and 2K output — the catch is a research-only license. Plus: Z.ai's ZCode data-exfil fallout.
Copy markdown
Transparency and 10-reference editing, in the open
Qwen-Image-2.1 is a 7.12B DiT (32 layers, a Qwen3-VL 8B text encoder) that generates and edits in one model — with native RGBA transparency (no separate matting step) and composition from up to 10 reference images at native 2K. Weights are on Hugging Face and ModelScope now.
The catch: a research-only license
Unlike every prior Qwen-Image release, which shipped under Apache 2.0, 2.1 lands under the Qwen Research License — “research or evaluation purposes only,” with any revenue-generating use (even indirect) requiring a separate paid license you have to email for. Open weights you can study, not ship.
Build this weekend: a local transparent-asset pipeline
Because it emits true RGBA and takes up to 10 references, you can wire up a product-shot compositor or a sticker/icon generator that outputs drop-in PNGs — no background-removal step, no per-image API bill — on a single GPU. Just keep it to a research demo until you clear the license.
Elsewhere: Z.ai open-sources ZCode after an exfil scandal
After a Sept 18 teardown showed the ZCode agent silently building a 313MB encrypted archive (42,411 files, 86.6% Git history) and making 564 upload attempts — with Z.ai holding the only decryption key — the firm open-sourced ZCode under Apache 2.0 and apologized. You can now audit the client, though the source can't prove server-side deletion.