We are excited to announce the integration of Nunchaku 4-bit diffusion inference into the Diffusers library. This new capability enables efficient, high-quality text-to-image generation with significantly reduced memory and computational overhead, making state-of-the-art diffusion models more accessible on consumer hardware.
What is Nunchaku?
Nunchaku is a lightweight inference framework that leverages 4-bit quantization (nvfp4) to compress large diffusion models. By reducing the precision of weights and activations, Nunchaku dramatically lowers memory usage and speeds up inference while preserving image quality.
Available Model
A ready-to-use model, OzzyGT/Krea2Turbonunchakulite_nvfp4, is now available on the Hugging Face Hub. This 8 billion parameter text-to-image model has been optimized with Nunchaku's 4-bit quantization, offering a strong balance between quality and efficiency.
- Model Name: OzzyGT/Krea2Turbonunchakulite_nvfp4
- Task: Text-to-Image
- Parameters: 8B
- Last Updated: July 8, 2026 (15 days ago)
- Downloads: 51
- Likes: 2
How to Use
To use the Nunchaku-optimized model with Diffusers, simply load it from the Hub and run inference as usual. The quantization is handled automatically, so you get the benefits of 4-bit inference without any extra code changes.
from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained("OzzyGT/Krea_2_Turbo_nunchaku_lite_nvfp4")
image = pipe("a beautiful landscape").images[0]
image.save("output.png")
Benefits
- Reduced Memory Footprint: 4-bit quantization cuts model size by up to 4x compared to full-precision versions.
- Faster Inference: Lower precision allows for faster computation on compatible hardware (GPUs with FP4 support).
- Preserved Quality: Despite aggressive quantization, image quality remains high, suitable for production and creative use.
Looking Ahead
As of mid-2026, Nunchaku represents a key step toward democratizing diffusion models. With ongoing hardware support for low-precision arithmetic, we expect 4-bit inference to become the new standard for efficient generative AI.
Try the model today and share your results with the community!