Zyphra Releases ZAYA1-8B: A Reasoning MoE Trained on AMD Hardware That Punches Far Above Its Weight Class

Zyphra released ZAYA1-8B, an 8.4B-parameter MoE reasoning model with only 760M active parameters.

The model was trained on AMD Instinct MI300 hardware and optimized with MoE++ and Markovian RSA to cut compute and memory costs.

ZAYA1-8B is open-sourced on Hugging Face under Apache 2.0 and also available as a Zyphra Cloud serverless endpoint.

Its math and coding benchmarks reportedly rival or beat several much larger frontier models, including DeepSeek, Gemini, Claude, and GPT-5 variants.