Zyphra ZAYA1-8B-Diffusion-Preview Review — First MoE Diffusion LLM Converted From Autoregressive

Zyphra ZAYA1-8B-Diffusion-Preview (May 14, 2026) converts the ZAYA1-8B autoregressive reasoning model into a discrete diffusion LM using the TiDAR recipe. First MoE diffusion model in the field. Generates 16 tokens per forward pass. Claims 4.6x (lossless) to 7.7x (logit-mixing) inference speedup. Trained on AMD Instinct MI300X. Preview-stage: no RL training applied, pass@ evals only, weights not yet publicly confirmed. Rating: 3/5.