Zyphra Releases Open EEG Foundation Model ZUNA1.1

ZUNA1.1 is an EEG restoration-specialized foundation model released by Zyphra, supporting noisy channel removal, denoising, and low-density-to-high-density electrode upsampling.

The key differentiator is its position-aware diffusion autoencoder architecture, which represents channels as 3D scalp coordinates rather than array indices. This allows it to handle arbitrary electrode layouts, from 4-channel consumer headbands to 256-channel research caps, and even generate signals for positions not seen in the training data.

Changes from ZUNA1: The architecture remains the same, but the training approach has been substantially improved to handle messy real-world data.

Comparison with existing methods:

• MNE spherical spline interpolation: performance drops sharply when many channels are missing or electrode density is low
• Fixed-channel CNN/MAE models: applicable only to specific channel configurations, lacking generality
• ZUNA1.1 shows improved restoration quality over both baselines, both quantitatively and qualitatively

The model weights are released under the Apache 2.0 license, enabling free commercial use, and technical details are described in the ZUNA1 arXiv paper (2602.18478).