Switch language한국어
Back to the list

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

TL;DR AI

Key summary

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

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

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

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

Read the original