Zyphra's ZAYA1-8B Matches Frontier AI Models on Benchmarks Using Under 1 Billion Active Parameters, Trained on AMD Hardware

Summary

Zyphra released ZAYA1-8B, a model that matches or competes with frontier AI models like DeepSeek-R1, Claude Sonnet 4.5, and Gemini 2.5 Pro on math, reasoning, and coding benchmarks — while using less than 1 billion active parameters (760M). It was trained entirely on AMD hardware, a notable departure from the NVIDIA-dominated AI training ecosystem. The model uses a Markovian RSA inference method that allows performance to scale with compute budget. However, it underperforms on agentic tasks like tool calling and multi-step instruction following. The article highlights this as significant for researchers focused on test-time compute methods and for those interested in AMD's potential in the AI hardware space.