Zyphra Releases ZUNA1.1 EEG Foundation Model with Variable Input Lengths

Zyphra has announced the release of ZUNA1.1, an updated electroencephalography (EEG) foundation model. Launched on July 16, 2026, and made available under the Apache 2.0 license, this new version introduces significant flexibility with variable-length inputs ranging from 0.5 to 30 seconds. This enhancement addresses a key limitation of its predecessor, ZUNA1, which was restricted to fixed five-second segments.

The ZUNA1.1 model functions as a 380-million-parameter masked diffusion autoencoder. It specializes in reconstructing, denoising, and upsampling scalp-EEG data across diverse channel layouts. This capability is crucial for handling the inherent variability and noise in real-world EEG recordings, making it a valuable tool for researchers and developers in neurotechnology and AI.

Key facts:

Feature	Detail
Model Name	ZUNA1.1
Release Date	July 16, 2026
License	Apache 2.0
Input Lengths	Variable, 0.5 to 30 seconds
Parameters	380 million

Enhanced Flexibility for EEG Data

The primary advancement in ZUNA1.1 is its ability to process EEG signals of varying durations. Unlike ZUNA1’s fixed five-second window, ZUNA1.1 can accommodate inputs from 0.5 to 30 seconds. This adaptability is vital because actual EEG recordings often differ in length, and data quality can fluctuate within sessions due to noisy channels or dropouts. The model’s design allows it to work with arbitrary electrode layouts, from simple four-electrode headbands to complex 256-channel research caps.

ZUNA1.1’s core function involves denoising existing EEG segments and channels, reconstructing missing data, and predicting novel channel signals based on physical coordinates on the scalp. This is achieved through a unique tokenization process where each channel is sliced into 0.125-second segments, each becoming a continuous-valued token. A 4D rotary positional encoding (x, y, z, t) for each token ensures channel-agnostic processing, enabling the model to generate signals at unrecorded positions.

Architectural and Training Improvements

While the parameter count remains at 380 million, ZUNA1.1 incorporates architectural and training enhancements focused on stability and robustness. The model is built as a transformer encoder-decoder diffusion autoencoder. It runs efficiently on consumer GPUs and can even perform acceptably on CPUs for many workloads.

Improvements in ZUNA1.1 include a richer mixture of reconstruction tasks during training. While ZUNA1 trained on uniformly random whole-channel dropout, ZUNA1.1 trains on four distinct dropout patterns. These patterns simulate various real-world data loss scenarios, such as whole-channel dropouts, removal of short time stretches across all channels, clustered gaps in space and time, and scattered missing individual data points. This diverse training regimen enhances the model’s ability to handle complex data imperfections.

Data Corpus and Performance

Zyphra has also expanded the training data corpus for ZUNA1.1. By implementing a per-channel, per-second quality scoring system, the team was able to process more usable signal, growing the corpus from approximately 2 million to 3.5 million channel-hours of public EEG data. This refined preprocessing strategy, along with the precomputation of two filter variants per recording (0.1–45 Hz bandpass and 0.01 Hz highpass with notch), contributes to the model’s improved generalization capabilities.

Despite the increased flexibility, ZUNA1.1 maintains or improves its reconstruction Normalized Mean Squared Error (NMSE) compared to ZUNA1 on held-out tasks. Both models significantly outperform classical spherical-spline interpolation. A region-based test, which involves reconstructing electrodes from one brain region using data from others, showed ZUNA1.1 outperforming both ZUNA1 and spherical-spline methods, indicating its practical utility in more realistic scenarios.

Implications for Indian AI and Neurotech Research

For the Indian AI and neurotech community, the open-source release of ZUNA1.1 under the Apache 2.0 license is significant. It provides researchers and startups with a powerful, flexible tool for analyzing complex EEG data without proprietary restrictions. This could accelerate developments in brain-computer interfaces, neurological disorder diagnostics, and cognitive science research within India, fostering innovation in areas like affordable healthcare solutions and assistive technologies. The availability of weights on Hugging Face and inference code on GitHub lowers the barrier to entry for experimentation and application development.

Source: MarkTechPost, https://www.marktechpost.com/2026/07/17/zyphra-releases-zuna1-1-an-apache-2-0-eeg-foundation-model-with-variable-length-inputs-from-0-5-to-30-seconds/