ZAYA1-8B debuts with AMD-trained MoE

This is a real hardware-and-architecture story, not just another benchmark post: Zyphra is trying to prove that careful MoE design plus AMD-scale infrastructure can buy frontier-ish quality density.

• –ZAYA1-8B is small in active parameters but large in total capacity, which makes it interesting for latency and cost-sensitive deployments
• –The AMD/IBM training setup is part of the headline; the release is also a validation of AMD as a serious large-scale training platform
• –Zyphra is leaning on CCA, a new router, and Markovian RSA, so the model claim is as much systems work as model work
• –Apache 2.0 weights on Hugging Face make it usable for teams that want to study or adapt the stack instead of just reading the paper
• –The benchmark comparisons are ambitious, but the practical test will be whether developers can reproduce the gains outside Zyphra’s own serving setup