ZUNA1.1

Electroencephalography data is rarely clean. Electrodes come loose, movement artifacts wipe out stretches of signal, and labs record with wildly different montages — a 4-channel consumer headband, a clinical 10-20 cap, a 256-channel research array. The standard repair tool is spherical spline interpolation, ubiquitously deployed in the MNE package, which estimates a missing electrode from the spatial smoothness of its neighbors and has no notion of what EEG looks like in time.

ZUNA1.1, released in July 2026 by Zyphra, attacks that problem generatively. It is a 380M-parameter transformer encoder-decoder diffusion autoencoder trained to reconstruct masked scalp EEG: given whatever channels and time spans survive, it fills in the rest. The design choice that makes it flexible is positional: every token carries a 4D rotary positional encoding over the electrode's 3D scalp coordinate plus a time index, so the model is told where a channel sits rather than which array slot it occupies. That makes it montage-agnostic without retraining, and lets it synthesize signal at positions that were never recorded.

This distinguishes ZUNA1.1 from most EEG foundation models. LaBraM, CBraMod, CSBrain, and LUNA learn transferable representations for downstream decoding; ZUNA1.1 targets the raw waveform itself, producing signal rather than embeddings. It iterates on Zyphra's earlier ZUNA1, keeping that model's architecture and parameter count while training for far greater flexibility.

#Key Features

• Montage-agnostic conditioning: Because each token is positioned by its electrode's (x, y, z) location rather than a channel index, the same weights serve a 4-channel headband or a 256-channel research cap with no reconfiguration.
• Variable-length windows: Segment lengths from 0.5 to 30 seconds are sampled on a 0.125 s token grid during training, replacing ZUNA1's fixed 5-second windows.
• Four reconstruction regimes: Training mixes whole-channel dropout, full-time dropout across all channels, clustered channel-time gaps, and scattered random-uniform dropout, covering the ways real recordings get corrupted.
• Spatial super-resolution: Sparse montages are upsampled by requesting new channels by name, or a target channel count placed away from existing electrodes.
• Open weights under Apache-2.0: Weights, inference code, and a zuna pip package are released permissively, and inference needs under 1 GB of VRAM.

#Technical Details

ZUNA1.1 uses 16 encoder and 16 decoder layers with a model dimension of 1024 and a head dimension of 64. Each channel is sliced into 32-sample segments (0.125 s at 256 Hz), turned into continuous-valued tokens, and serialized in channel-by-time order; the 4D rotary encoding discretizes each spatial axis into 100 bins. The encoder compresses the input into a latent that conditions the decoder through adaptive RMS normalization, and the decoder is trained with a rectified-flow objective plus an auxiliary MMD loss on the encoder latents. Training used roughly 3.5 million channel-hours of public EEG, a 75% increase over ZUNA1's ~2M, made possible by a per-channel, per-second quality score that salvages signal from partially noisy recordings; two filter variants per recording teach robustness to heterogeneous preprocessing.

On held-out evaluations spanning ANPHY-Sleep, BerlinBCI, BCI2000, and AAD, ZUNA1.1 matches or improves on ZUNA1's normalized mean squared error while handling a far wider task distribution, and both clearly outperform spherical spline interpolation, with the margin widening as the fraction of missing channels grows. A harder region-occlusion test — deleting every electrode over one brain region and reconstructing it from the remaining seven — also favors ZUNA1.1. Early ablations at roughly 1.2B parameters did not meaningfully improve reconstruction or downstream performance, so the release stayed at 380M.

#Applications

ZUNA1.1 slots into EEG preprocessing pipelines as a repair step: reconstructing dead or dropped electrodes, cleaning motion-corrupted time spans on selected channels, harmonizing datasets collected on incompatible montages, and upsampling low-density consumer or clinical recordings toward denser layouts. It operates directly on MNE .fif files, taking the union of the file's own bad-channel flags and BAD_ annotations plus any channels or time spans the user names, and Zyphra hosts a browser playground for no-code trials. The authors are explicit that this is a research tool, not a diagnostic one: reconstructed channels are imputed data, and ZUNA1.1 can produce plausible signal that was never present.

#Impact

ZUNA1.1 reframes what an EEG foundation model is for, treating signal restoration rather than representation learning as the primary objective, and makes a concrete case that a learned generative prior beats the interpolation defaults the field has relied on for decades. The permissive Apache-2.0 release of weights and inference code, plus a pip-installable package that integrates with MNE, removes most of the friction that has kept EEG foundation models confined to their authors' benchmarks. Two caveats matter: the work is a preprint awaiting peer review, and the authors report a growing dissociation during training between reconstruction accuracy and the usefulness of the learned representations for downstream classification, so strong reconstruction is not evidence of a general-purpose EEG encoder.