Qwen Image 2.1 Turbo
Qwen's fast version of Qwen-Image-2.1 generates and edits images in 8 steps instead of 40, with the same 7B model underneath. Here is what it does, how to run it and what the license allows.
How to run it
Three ways to use Qwen Image 2.1 Turbo
Pick by how much setup you want.
ComfyUI
Download qwen_image_2.1_turbo_bf16 or the int8 file from Comfy-Org/Qwen-Image-2.1, or load the Turbo LoRA on the base model. Use 8 steps and CFG 1.
GGUF on smaller GPUs
Community GGUF builds (for example unsloth/Qwen-Image-2.1-Turbo-GGUF) range from about 2.5 GB to 14 GB on disk; you also need the VAE and the Qwen3-VL text encoder.
Hosted API
Alibaba Cloud Model Studio serves qwen-image-2.1-turbo per image, which avoids the research-license limits on self-hosted weights.
What it is
What Qwen Image 2.1 Turbo changes
Qwen released Qwen-Image-2.1-Turbo on October 9, 2026, three weeks after Qwen-Image-2.1. It is an accelerated checkpoint of the same model, not a new architecture.
Speed
8 steps instead of 40
Turbo runs 8 denoising steps against the base model's 40, at CFG 1, and reuses text and reference-image context with a prefix KV cache. Qwen has not published a speed multiple or a VRAM figure for Turbo.
- Same 7B visual generation architecture
- Sampling schedule ships with the weights
- Changing num_inference_steps alone does not override it
Generate and edit
Text-to-image and image editing
Like Qwen-Image-2.1, Turbo covers text-to-image and editing. Qwen's showcase includes transparent (RGBA) images, poster typography, UI layouts and multi-reference composites.
- Native 2K sizes such as 2048 × 2048
- Transparent PNG output
- Multi-image reference edits
License
Non-commercial unless you ask Qwen
Both Qwen-Image-2.1 and the Turbo weights use the Qwen Research License, which is for non-commercial use only; commercial use needs separate permission from Qwen. The original Qwen-Image was Apache 2.0.
- Open weights on Hugging Face and ModelScope
- Non-commercial license
- Commercial use: hosted API or a separate agreement
Do not mix them up
Official Turbo vs the Viggle turbo
Qwen-Image-2.1-viggle-turbo is a separate LoRA adapter from Viggle, released about two weeks earlier with 4 to 9 step modes. The official Qwen Turbo is a full checkpoint. Speed and VRAM claims for the Viggle version do not apply to Qwen's.
- Viggle: third-party LoRA
- Qwen: official full weights
- Both build on Qwen-Image-2.1
Specs
Qwen Image 2.1 Turbo specs at a glance
From Qwen's model card, GitHub and Alibaba Cloud Model Studio, October 2026.
- Developer
- Qwen (Alibaba)
- Released
- October 9, 2026
- Weights
- Qwen/Qwen-Image-2.1-Turbo on Hugging Face and ModelScope
- Size
- 7B visual generation model (about 7.1B parameters, BF16)
- Steps
- 8 (base Qwen-Image-2.1: 40)
- Modes
- Text-to-image, image editing
- License
- Qwen Research License (non-commercial)
- Hosted API
- qwen-image-2.1-turbo on Model Studio, about $0.014–$0.016 per image
- ComfyUI / GGUF
- Comfy-Org bf16 and int8 files plus a Turbo LoRA; community GGUF from about 2.5 GB
- In LazyKiwi
- Not available; the image generator offers Qwen Image
Which Qwen Image
Qwen Image 2.1 Turbo vs 2.1 vs the original
Facts from Qwen's model cards and GitHub README.
| Model | Released | Steps / size | License |
|---|---|---|---|
| Qwen-Image-2.1-Turbo | Oct 9, 2026 | 8 steps, 7B | Non-commercial |
| Qwen-Image-2.1 | Sept 20, 2026 | 40 steps, 7B, up to 10 references | Non-commercial |
| Qwen-Image | Aug 4, 2025 | About 20B | Apache 2.0 |
FAQ
Qwen Image 2.1 Turbo FAQ
What is Qwen Image 2.1 Turbo?
Qwen-Image-2.1-Turbo is an accelerated checkpoint of Qwen-Image-2.1, released by Qwen on October 9, 2026. It generates and edits images in 8 denoising steps instead of 40, with the same 7B architecture.
Is Qwen Image 2.1 Turbo open source and free for commercial use?
The weights are public, but the Qwen Research License allows non-commercial use only. Commercial use needs permission from Qwen, or use the hosted Model Studio API.
Is there a Qwen Image 2.1 Turbo GGUF?
Not from Qwen. Community GGUF quantizations exist, such as unsloth/Qwen-Image-2.1-Turbo-GGUF, from Q2_K (about 2.5 GB) to F16 (about 14 GB).
Is there a ComfyUI workflow?
Comfy-Org publishes Turbo weights and a Turbo LoRA. Use the Qwen-Image-2.1 workflow with 8 steps and CFG 1.
Qwen Image 2.1 vs Turbo: what is the difference?
Same model family and architecture; Turbo uses 8 steps with a built-in schedule, while the base model uses 40. Qwen has not published side-by-side quality or speed numbers.
Can I use it on LazyKiwi?
Not at the moment. LazyKiwi's image generator offers Qwen Image, plus Z-Image, Nano Banana, GPT Image and Seedream.
How many steps does Qwen Image 2.1 Turbo use?
8 denoising steps, against 40 for the base Qwen-Image-2.1, with CFG 1 in the Comfy workflow. The 7B architecture is the same.
Can Qwen Image 2.1 Turbo edit images?
Yes. Like Qwen-Image-2.1 it covers text-to-image and image editing in one checkpoint.
How much VRAM or disk space does Qwen Image 2.1 Turbo need?
The BF16 weights are about 7.1B parameters. Community GGUF builds run from about 2.5 GB (Q2_K) to 14 GB (F16) on disk, which is the usual route on smaller GPUs; you also need the text encoder and VAE.
Where can I use Qwen Image 2.1 Turbo commercially?
The self-hosted weights are under the non-commercial Qwen Research License. For commercial work, Alibaba Cloud Model Studio serves qwen-image-2.1-turbo as a paid API at about $0.014–$0.016 per image.
Try Qwen Image for free, no ComfyUI
Try Qwen Image in LazyKiwi's image generator, in the browser.
More models
Image models you can run in LazyKiwi today.