Qwen-Image-2.1-Turbo weights are out, under a non-commercial license

On 9 October 2026, Qwen published the Qwen-Image-2.1-Turbo checkpoint and, in the same GitHub news list, said the Qwen-Image-2.1 Pro and Turbo APIs are officially live on Alibaba Cloud Model Studio. The base model is older. That list dates Qwen-Image-2.1 itself to 20 September 2026. Turbo is an accelerated checkpoint of that model, not a new architecture.

The Qwen account posted at 13:24 UTC the same day. The post calls the files open weights, names the eight-step checkpoint and the same 7B architecture, and says the Pro and Turbo APIs are live. It also asserts that fewer steps do not lower quality. The model card gives no quality comparison to back that up.

What the card says the checkpoint is

The card calls Turbo an accelerated checkpoint of Qwen-Image-2.1 for text-to-image generation and image editing in 8 denoising steps. It uses the same 7B visual generation architecture and loads with QwenImage21Pipeline in Diffusers. The Hugging Face model record lists 7,115,124,736 BF16 parameters, createdAt 04:50 UTC on 9 October 2026, and base_model Qwen/Qwen-Image-2.1. The card’s license fields are license: other and license_name: qwen-research.

The checkpoint includes its recommended sampling schedule, so the card says you do not configure the scheduler by hand. Generation uses CFG=1 by default. Prefix KV caching reuses the text and reference-image context across denoising steps. The card’s examples pass use_kv_cache=True.

The saved 8-step schedule loads automatically. Setting num_inference_steps alone does not override it. An explicit sigmas argument at call time does, and the card says other schedules have not been evaluated for this checkpoint. That is the practical point of keeping the sampler that belongs to the checkpoint: Qwen-Image is one of the models where a swapped scheduler is a compatibility question, and this card says the Turbo schedule is already stored with the weights.

The card’s text-to-image example uses 1680 by 2512. The editing example uses 2048 by 2048. It also lists the same resolution presets as Qwen-Image-2.1: 2048 by 2048, 2400 by 1792, 1792 by 2400, 2528 by 1696, 1696 by 2528, 2752 by 1536, and 1536 by 2752.

Installation, as the card writes it, is a CUDA-compatible PyTorch, then Diffusers installed from its Git repository, plus transformers>=5.17.0, accelerate, and pillow. The card says the checkpoint needs Diffusers support for pipeline-configured sampling sigmas, added in pull request 14950.

The README links Pro and Turbo console URLs and calls both APIs officially live.

The Qwen Research License

The weights are open weights under the Qwen Research License. The file is headed “Qwen RESEARCH LICENSE AGREEMENT” and gives its release date as 20 September 2026. “We” is defined as Hangzhou Tongyi Laboratory Technology Co., Ltd. The agreement does not use the name Alibaba; the release was announced by the Alibaba Qwen account.

Section 2 grants a non-exclusive, worldwide, non-transferable, royalty-free limited license to use, reproduce, distribute, copy, create derivative works of, and modify the materials “FOR NON-COMMERCIAL PURPOSES ONLY.” The next sentence is: “You shall not use the Materials for any commercial purpose without obtaining a separate commercial license from us.” Commercial requests go to the address in that section.

The GitHub introduction says “We are excited to open-source Qwen-Image-2.1.” That sentence is the repository’s. The license text is the restriction that applies to the weights.