Qwen-Image-2.1: transparent-background image gen on a 16GB GPU

The open model surging up Hugging Face outputs real alpha channels and native 2K in one checkpoint — but its license quietly flipped to research-only.

Nowline OCT 1 1:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • RGBA output, no matting step

    Qwen-Image-2.1 is a 7B model that generates images with a true alpha channel — transparent cutouts and layered edits straight from a prompt, plus native 2K (up to 2752px) and a single checkpoint for both generation and editing with up to 10 reference images. It's the top open image model on Hugging Face this week.

  • Runs on ~16GB, ~5-25s an image

    At Q8 it fits in ~16-17GB of VRAM, or down to ~3GB with CPU offload, and renders a 1MP image in 5-25 seconds on a modern GPU. It's already wired into ComfyUI, Diffusers, vLLM-Omni and SGLang — a real weekend build for product mockups, sticker and asset pipelines, or transparent UI art with no Photoshop round-trip.

  • The license catch: research-only now

    Heads up before you build on it: the original Qwen-Image was Apache 2.0, but 2.1 moved to the Qwen Research License — non-commercial only, with commercial use requiring a separate agreement with Alibaba. It's "weights-available," not open source, so it's fine for prototypes but a blocker for anything you ship.

  • Need commercial? Reach for Qwen3.8-27B

    If you want weights you can actually deploy, Qwen3.8-27B (also trending) stays Apache 2.0: a 27B vision-language MoE with 262K context, extendable to 1M, scoring 90.3% on LiveCodeBench and 84.3% on OSWorld computer-use. Self-host it or call it through OpenRouter.