Diffusion language model "ZAYA1-8B-Diffusion-Preview" developed with AMD AI chips debuts, converting autoregressive models into diffusion models

Zyphra unveiled ZAYA1-8B-Diffusion-Preview, the first diffusion language model trained on AMD hardware.

The model was converted from the pretrained ZAYA1-8B autoregressive LLM, showing a practical path from LLMs to diffusion-based generation.

Zyphra says the preview model decodes 4.6x to 7.7x faster while largely preserving benchmark quality.

The result highlights a way to reduce inference bottlenecks, especially memory-bandwidth limits, in language generation.