Qwen-Image-2.1: open 7B with transparency, but no commercial use
Alibaba's 7B matches paid tools and runs local, but weights are research-only. Also: MiMo-V2.6-Pro tops the open board and Rabbit's OS3 goes multi-platform.
Copy markdown
Transparent PNGs with no cut-out step
Qwen-Image-2.1 generates and edits in native RGBA, so product shots, logos, and stickers come out with a real alpha channel — skip background removal entirely. It takes up to 10 reference images at once for group scenes or virtual try-ons, and edits from circles, masks, or painted marks.
It fits on a single 3090
The model is 7B and the weights are up on Hugging Face, GitHub, and ModelScope with a live demo, running on one consumer GPU. Alibaba says it beats most closed image models — on its own benchmark, with independent numbers still pending.
The catch: you can't ship what it makes
The weights are open but the license is research-only; commercial use needs a separate paid license from Qwen. Perfect for prototyping a design tool this weekend, a blocker the moment its output goes into something you sell.
Update: MiMo-V2.6-Pro tops the open-weights board
The Pro sibling to yesterday's Flash is now the #1 open-weights model on the Artificial Analysis Index at 46 — a 1.02T/42B MoE, MIT-licensed, 1M context, omnimodal, at $0.44/$0.87 per M tokens. Xiaomi also open-sourced 7,000+ RL environments and its training framework.
Rabbit's OS3 leaves the R1 behind
Rabbit's agentic OS now runs on Windows, Mac, Linux, and cloud, driving your real apps by navigating their interfaces instead of via integrations. There's no developer API or independent benchmark yet, so weigh the demos carefully.