# Use Qwen-Image 2.1 Turbo locally without ComfyUI — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/bd01d620a820cd3a

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-10-11 12:28 ingested 2026-10-12 02:38

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1x36zdx/use_qwenimage_21_turbo_locally_without_comfyui/)

## retained text



```
I'm the developer of [Quartermaster](https://quartermaster-labs.github.io/Quartermaster/), a free, open-source (MIT) local inference server. It drives stable-diffusion.cpp and also serves LLMs, TTS and speech-to-text from the same app, swapping models in and out of VRAM as you use them.  
  
Models download from inside the app: search Hugging Face, pick a quant, and Quartermaster writes the configuration itself. It matches each model to its VAE and text encoders and picks the right steps and cfg. Qwen-Image 2.1 Turbo comes up at 8 steps and cfg 1, and real transparent PNGs work too: tick "transparent" and it adds the exact prompt wording the RGBA VAE needs. 

To run 2.1 Turbo, grab these three (suggested files) from the built-in model browser and you're done:

* the DiT: [Abiray/Qwen-Image-2.1-Turbo-GGUF](https://huggingface.co/Abiray/Qwen-Image-2.1-Turbo-GGUF) (I use Q8\_0)
* the 2.1 VAE: [qwen\_image\_2.1\_vae\_bf16.safetensors](https://huggingface.co/Comfy-Org/Qwen-Image-2.1/blob/main/vae/qwen_image_2.1_vae_bf16.safetensors).
* [Qwen3-VL-8B-Instruct GGUF](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct-GGUF) plus its mmproj file
```



document fetched 2026-10-12 00:29 · canonical https://reddit.com/r/StableDiffusion/comments/1x36zdx/use_qwenimage_21_turbo_locally_without_comfyui
