Qwen-Image-2.1-Turbo-Quantized: Run Qwen-Image 2.1 Turbo on 4GB VRAM

This repository distributes the Qwen-Image-2.1-Turbo model converted into various quantization formats. A total of 11 files are provided, including GGUF, FP8, NVFP4, and MLX, allowing selection based on hardware specifications. In particular, it includes Q4_K_M GGUF and MLX 4-bit files that operate on 4GB-class VRAM, supporting image generation in low-spec environments.

Unlike conventional methods, this model does not apply built-in safety checkers or content filters. It generates images directly from the original prompt without additional configuration, using 8-step denoising and CFG 1.0 as default values to enable fast inference. Generating a 1024×1024 resolution image in a GPU environment takes approximately 10–20 seconds.

It is optimized for compatibility with major runtimes such as ComfyUI, stable-diffusion.cpp, and MFLUX. In ComfyUI, it is used with the official workflow template, while in the terminal, images can be generated using CLI commands via stable-diffusion.cpp. In MLX environments, it runs at 4/6/8-bit precision on Apple Silicon and Linux with CUDA 13 support.

This is suitable for developers looking to set up Qwen-Image 2.1 Turbo locally on low-spec GPUs or Apple Silicon Macs. Text encoder and VAE files are also provided, allowing immediate use without additional downloads. It is distributed under the Qwen Research License, maintaining the original model's architecture and generation quality while reducing only memory usage.