Qwen-Image-2.1-Turbo is an accelerated checkpoint of Qwen-Image-2.1 for text-to-image generation and image editing. It uses eight denoising steps and the same 7-billion-parameter visual generation architecture.[1] The Qwen team released it on October 9, 2026, alongside hosted Pro and Turbo APIs.[2]

Model specifications

Property	Specification
Visual generation component	7B parameters[1]
Recommended sampling	8 denoising steps[1]
Diffusers pipeline	`QwenImage21Pipeline`[1]
Default classifier-free guidance	CFG = 1[1]
Downloadable checkpoint	`Qwen/Qwen-Image-2.1-Turbo` on Hugging Face[1]
Hosted model identifier	`qwen-image-2.1-turbo`[5]
Model license	Qwen Research License Agreement[4]

The model card demonstrates transparent-image generation, typography, single-image editing, and composition from multiple reference images. Its examples use 2048 x 2048 editing and 1680 x 2512 text-to-image generation.[1]

Sampling and local execution

Turbo saves its sampling schedule with the checkpoint. KV caching reuses text and reference-image context across denoising steps.[1]

Diffusers pull request #14950 added the pipeline configuration field sample_sigmas. The pipeline passes the selected schedule to the scheduler, and the schedule determines the step count. An explicit sigmas argument overrides the saved schedule; setting num_inference_steps alone does not. The scheduler implementation itself was not changed by this addition.[3]

The model card directs users to install a CUDA-compatible PyTorch build, current Diffusers source with that change, transformers>=5.17.0, Accelerate, and Pillow. Its examples load the pipeline in BF16 on CUDA. Qwen states that alternative sampling schedules have not been evaluated for this checkpoint.[1]

Hosted access

Alibaba Cloud Model Studio provides hosted inference under qwen-image-2.1-turbo. Its model information page lists text and image input, image output, and fine-tuning support. Function calling, structured outputs, and web search are unsupported.[5] Qwen also distributes the checkpoint for execution in users' own environments.[2]

License

The Qwen Research License Agreement grants use, modification, and redistribution of the model materials for non-commercial research or evaluation. Commercial use requires a separate license from the licensor. Redistribution requires the agreement, notices identifying modified files, and the specified attribution notice. The agreement also imposes attribution conditions when the materials or their outputs are used to create or improve a distributed AI model.[4]

References

Qwen. Qwen-Image-2.1-Turbo model card. Hugging Face. Accessed October 10, 2026.
Qwen. Qwen-Image-2.1 repository, News, Hosted APIs, and Turbo quick start. GitHub. October 9, 2026 release entry. Accessed October 10, 2026.
Hugging Face Diffusers. Configure fixed sampling sigmas on QwenImage21Pipeline, pull request #14950. Merged October 5, 2026. Accessed October 10, 2026.
Hangzhou Tongyi Laboratory Technology Co., Ltd. Qwen Research License Agreement. License release date September 20, 2026. Accessed October 10, 2026.
Alibaba Cloud Model Studio. qwen-image-2.1-turbo model information. Updated October 10, 2026. Accessed October 10, 2026.

Editing: Qwen-Image-2.1-Turbo