Zyphra Releases ZUNA1.1, an Open EEG Foundation Model for Brain Signals of Any Length

Zyphra has released ZUNA1.1, an openly licensed foundation model for electroencephalography (EEG) — the electrical recordings of brain activity — on July 17. Published under a permissive Apache 2.0 license, it is a bid to do for messy neural data what earlier foundation models did for text and images.

The model

ZUNA1.1 is a 380-million-parameter transformer encoder-decoder built as a diffusion autoencoder. Its headline upgrade over the original ZUNA1 is flexibility: it accepts variable-length inputs from 0.5 to 30 seconds, rather than being locked to fixed 5-second segments — a practical difference when real recordings do not arrive in tidy windows.

What it does

The model denoises EEG, reconstructs missing channels and upsamples recordings — filling in the gaps that plague real-world brain data. On held-out reconstruction error, Zyphra reports it matches or beats classical spherical-spline interpolation, the standard technique it aims to replace. A diffusion autoencoder learns to compress signals and then regenerate them, which is what lets the model infer plausible values for channels that dropped out or were never recorded in the first place.

From headbands to research rigs

Crucially, it generalizes across hardware. Using position-aware tokenization, ZUNA1.1 works across arbitrary electrode layouts — from 4-electrode consumer headbands up to 256-channel research rigs. That cross-device portability is exactly the fragmentation that has kept neural data siloed and hard to build on.

Open weights

Trained on roughly 3.5 million channel-hours of public EEG — up from about 2 million for the prior version — the model, its code and its weights are all freely available. For a field where labeled brain data is scarce and expensive, an open, hardware-agnostic base model lowers the barrier for startups and academics building neurotechnology and brain-computer interfaces. Zyphra, better known for efficient language and multimodal models, is applying the same efficiency-first philosophy to biosignals, betting that open infrastructure seeds the developer ecosystem that ultimately drives adoption.