> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Qwen releases Qwen-Image-2.1-Turbo, generating and editing images in 8 steps
- URL: https://globalfeed.ai/en/qwen-releases-qwen-image-2-1-turbo-generating-and-editing-images-in-8-steps/
- Published: 2026-10-10T18:55:21.000Z
- Updated: 2026-10-10T18:55:21.000Z
- Description: The Qwen team released Qwen-Image-2.1-Turbo, a faster version of its Qwen-Image-2.1 image generation and editing model. Turbo cuts the step count from the 40 used in the base model's examples to 8. The weights are on Hugging Face and ModelScope, and Pro and Turbo APIs went live on Alibaba Cloud the
- Author: GlobalFeed Editor
- Tags: Qwen, Qwen-Image-2.1-Turbo, image generation, image editing, x-Alibaba_Qwen, dil-en, elle

The **Qwen** team released **Qwen-Image-2.1-Turbo** on October 9, 2026, a faster version of its Qwen-Image-2.1 model for generating and editing images. According to the team's GitHub page, Turbo handles both text-to-image generation and editing of an existing image in 8 steps; the base model's example code uses 40\. The weights can be downloaded from Hugging Face and ModelScope. On the same day, Pro and Turbo versions of Qwen-Image-2.1 also went live as APIs on Alibaba Cloud Model Studio.

Diffusion models such as Qwen-Image build an image by starting from random noise and cleaning it up step by step, running the model again at each step. Going from 40 steps to 8 means an image can come out in far less time on the same hardware. Qwen has not published a time or speedup figure. The base Qwen-Image-2.1 model was released on September 20.

## Same architecture, with the step schedule built in

Turbo uses the base model's 7-billion-parameter image generation architecture; in the base model, a Qwen3-VL 8B encoder reads the prompt and the reference images. The recommended 8-step schedule ships with the checkpoint and Diffusers loads it automatically, so no manual setup is needed. According to the model card, changing the step count alone does not override this saved schedule, and Qwen says other schedules have not been evaluated for this checkpoint. Turbo also uses a prefix KV cache that reuses the prompt and reference image context across steps.

The examples on the model card show Turbo covering the same ground as the base model: portrait photography, human poses and motion, transparent (RGBA) images, typography and poster design, UI and infographic layouts, single-image transformations and combining several reference images into one scene. The base model accepts up to 10 reference images for editing. The default resolution is 2048x2048, with recommended sizes for seven aspect ratios including 16:9 and 9:16.

## Ways to run it, and the license

The model loads with the QwenImage21Pipeline in the Diffusers library, which requires Diffusers installed from source and transformers 5.17 or later. The model card also lists Draw Things and DiffusionBee, two Mac apps. No GPU memory figure is given for Turbo; for cards with limited memory, the GitHub page suggests CPU offload, which moves part of the model into system memory. The page does not say whether ComfyUI, vLLM-Omni, SGLang and LightX2V, which supported the base model from day one, already support Turbo.

Like the base model, Turbo is released under the Qwen Research License. The license allows using, modifying and sharing the model for research and evaluation; commercial use requires a separate commercial license from Qwen.

Qwen has not published quality comparisons with the base model or generation times for Turbo. Because faster variants can give different results from the base model, it is worth comparing the two on your own examples. For a fuller picture, follow evaluations from independent researchers as well.

Sources: [Hugging Face: Qwen-Image-2.1-Turbo model card](https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo?ref=globalfeed.ai), [GitHub: Qwen-Image-2.1](https://github.com/QwenLM/Qwen-Image-2.1?ref=globalfeed.ai), [ModelScope: Qwen-Image-2.1-Turbo](https://modelscope.cn/models/Qwen/Qwen-Image-2.1-Turbo?ref=globalfeed.ai).