Zyphra

Zyphra is a San Francisco- and London-based open AI lab founded in 2021, building hybrid SSM/Transformer foundation models co-designed with non-NVIDIA accelerators. Reached a $1B Series A valuation in October 2025 with backing from AMD Ventures, Intel Capital, IBM, Bison Ventures, Future Ventures, and others (~$110M total raised).

Architectural identity centers on the Zamba family — hybrid Mamba2 + shared-attention models in the 1.2B–7.4B parameter range, designed for low-latency on-device inference with KV-cache footprints up to 6x smaller than pure Transformers. The reasoning successor ZAYA1-8B (May 2026) is an 8B/700M-active MoE trained end-to-end on a full-stack AMD platform (MI300 GPUs), reportedly matching DeepSeek-R1-0528 on math/coding at well under 1B active. It is one of the most public foundation-scale demonstrations on AMD silicon to date — since scaled to a 74B/4B-active preview (also trained end-to-end on AMD). The hybrid line has also been extended to vision-language via Zamba2-VL (1.2B–7B).

Zyphra also publishes open training infrastructure (Zyda-2, a 5T-token pretraining dataset built with NVIDIA NeMo Curator) and is extending its hybrid-architecture toolkit into scientific domains via ZUNA, an EEG foundation model trained on 208 harmonized datasets.

Note: as of June 2026, Zyphra is still not tracked by Artificial Analysis — benchmark numbers on this page are self-reported from the technical reports.