Qwen Ships Image-2.1-Turbo as 8-Step 7B Research-License Model

Qwen released an accelerated checkpoint of its image model that gets to a finished picture in eight denoising steps.

Qwen-Image-2.1-Turbo, posted to Hugging Face, is "an accelerated checkpoint of Qwen-Image-2.1 for text-to-image generation and image editing with 8 denoising steps." The team notes it "uses the same 7B visual generation architecture and loads directly with QwenImage21Pipeline in Diffusers." Default inference runs at bfloat16 with CFG=1, and the checkpoint ships its own sampling schedule, "ready to use without manually configuring the scheduler." Setting num_inference_steps alone does not override it; only an explicit sigmas argument does.

From the model card: "prefix KV caching reuses the text and reference-image context across denoising steps." That's the trick that cuts repeated work when a run conditions on reference images.

The listed capabilities span text-to-image, image editing, transparent (RGBA) generation, portrait photography, human poses and motion, typography and poster design, UI and information layout, single-image transformation, and multi-reference composition. Supported aspect ratios top out at 2752x1536 for 16:9.

The catch is the license. The weights are released under the Qwen Research License Agreement, not a permissive commercial license. For readers tracking the steady drumbeat of Chinese open-weights image releases, our Alibaba tracker has logged 80 prior stories in 90 days; this one lands squarely as a research deliverable.

Only the saved 8-step schedule has been evaluated. In the card's own words, "other schedules have not been evaluated for this checkpoint."