Zyphra

Zyphra is a San Francisco-based AI research lab focused on efficient mixture-of-experts (MoE) language models with custom attention mechanisms designed to preserve reasoning quality at small active-parameter budgets. The company's flagship release as of May 2026 is ZAYA1-8B — an open-weight MoE model with 8.4 billion total parameters and only 760 million active per inference token.

On May 6, 2026, Zyphra published ZAYA1-8B with two notable claims: first, the model matches DeepSeek-R1 on AIME 2026 (89.1) and HMMT (71.6) and stays competitive with Claude Sonnet 4.5 on broader reasoning despite the much smaller active-parameter footprint. Second, the model was trained entirely on a 1,024-node AMD Instinct MI300X cluster built with IBM, using AMD Pensando Pollara networking — making it the first public frontier-quality math model trained outside the NVIDIA CUDA stack. The training claim sits alongside Cerebras and Groq as evidence that the non-NVIDIA frontier-AI training market is maturing.

ZAYA1-8B weights ship under the Apache 2.0 license — the most permissive open-model license, with no restrictions on commercial or competitive use. Local deployment requires Zyphra's vLLM fork (the upstream vLLM does not yet have the MoE routing and specialized attention patches Zyphra needs); serverless deployment via Zyphra Cloud avoids the runtime switch.