Alibaba's Qwen-Image-2.1-Turbo Hits 2.5M Downloads After Community Shrinks It
Community quant repack of Qwen-Image-2.1-Turbo hits 2.5M downloads and 4.5k likes on Hugging Face. Eleven formats: GGUF (Q4 to Q8), FP8, INT8 ConvRot, NVFP4 safetensors, and MLX 4/6/8-bit for Apple Silicon. Base model generates in 8 denoising steps at CFG 1, versus 40 steps for Qwen-Image-2.1.
Takeaways
- Community quant repack of Qwen-Image-2.1-Turbo hits 2.5M downloads and 4.5k likes on Hugging Face.
- Eleven formats: GGUF (Q4 to Q8), FP8, INT8 ConvRot, NVFP4 safetensors, and MLX 4/6/8-bit for Apple Silicon.
- Base model generates in 8 denoising steps at CFG 1, versus 40 steps for Qwen-Image-2.1.
- 1024x1024 generation takes 10 to 20 seconds on GPU after warmup; smallest quant is 4.00 GB.
- Ships ComfyUI workflow, stable-diffusion.cpp CLI recipe, and MFLUX setup for MLX inference.
- Licensed under Qwen Research License; no built-in safety filter, which drives part of its traffic.
A third-party quantization of Alibaba’s Qwen-Image-2.1-Turbo has recorded more than 2.5 million downloads on Hugging Face. The quantized repository converts the official BF16 checkpoint into 11 smaller builds for stable-diffusion.cpp, ComfyUI, and Apple Silicon, extending the model to hardware that cannot comfortably run the original release.
Hugging Face’s counter measures repository downloads rather than unique users or confirmed installations. The package has also ranked high on the platform’s trending list, giving the community conversion unusual visibility for an image-model repack.
Eight steps reshape the runtime
Qwen-Image-2.1-Turbo is Alibaba’s accelerated checkpoint of its 7-billion-parameter image model. It retains the architecture and advertised tasks of Qwen-Image-2.1, including text-to-image generation, instruction-based editing, RGBA transparency, and subject extraction.
The Turbo checkpoint generates images with eight denoising steps, compared with the base model’s default of 40. Denoising steps progressively turn random noise into an image, so reducing the count cuts the model’s main compute loop by a factor of five, although fixed overhead prevents a guaranteed fivefold reduction in total latency.
Classifier-free guidance, or CFG, is fixed at 1.0 for the recommended workflow. At that setting, negative conditioning is effectively unused, so negative prompts do not influence the result.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
