Zyphra ZAYA1-8B Review — MoE++ Reasoning Model, 760M Active Params, Beats Frontier Math on AMD

Zyphra ZAYA1-8B (May 6, 2026) is an 8.4B-parameter Mixture-of-Experts reasoning model with only 760M active parameters per token. MoE++ architecture with Compressed Convolutional Attention (CCA) delivers 8x KV-cache compression. Trained on AMD Instinct MI300X — no NVIDIA. AIME 2026: 89.1. LiveCodeBench: 65.8. Apache 2.0. Free weights. Rating: 4/5.