Zyphra

2026•Zyphra

ZAYA1-8B v1.8B

ZAYA1-8B is a compact 8.4B parameter MoE reasoning model from Zyphra, trained entirely on 1,024 AMD MI300X GPUs. It achieves frontier-level reasoning, mathematics, and coding performance competitive with models many times its size, released under Apache 2.0.

8.4B MoE model achieving frontier reasoning, math, and coding competitive with much larger models

Pretrained entirely on AMD MI300X hardware (1,024 GPUs) — no NVIDIA dependency

Open-source under Apache 2.0 license

PricingFree (Open Source, Apache 2.0)

Version1.8B

2026•Zyphra

ZUNA v1.0

Open Source

LLMs

ZUNA is Zyphra's first foundation model trained on brain data, aiming to advance noninvasive thought-to-text communication through brain–computer interfaces (BCIs). This 380M-parameter diffusion autoencoder model enhances the quality and usability of electroencephalography (EEG) data by reconstructing, denoising, and upsampling signals across various channel layouts.

Denoises existing EEG channels

Reconstructs missing EEG channels

Predicts novel channel signals based on scalp coordinates

PricingFree

Version1.0

2026•Zyphra

Online Vector Quantized Attention (OVQ) v1.0

Open Source

LLM

OVQ is a novel sequence mixing layer developed by Zyphra that aims to balance memory-compute costs and long-context capabilities more effectively than standard sequence mixing layers. It utilizes a sparse memory update to significantly increase memory state size while maintaining efficient training and inference characteristics.

Linear compute and constant memory complexity

Sparse memory update for increased memory capacity

Competitive performance on long-context tasks

Version1.0

RegionUnited States