Zyphra Releases ZAYA1-8B-Diffusion-Preview: The First MoE Diffusion Model Converted From an Autoregressive LLM With Up to 7.7x Speedup

Zyphra released ZAYA1-8B-Diffusion-Preview, a diffusion language model converted from its ZAYA1-8B autoregressive MoE checkpoint.

The model was further improved with mid-training and fine-tuning, and Zyphra says it keeps performance close to the original base model.

On AMD hardware, the preview reportedly delivers up to 7.7x faster inference, highlighting a strong efficiency gain.

The release suggests a practical way to repurpose existing LLMs into diffusion models without training from scratch.