Zyphra Releases ZUNA1.1: Apache 2.0 EEG Base Model With Variable-Duration Inputs From 0.5 To 30 Seconds
This week, Zyphra was released ZUNA1.1 under the Apache 2.0 license. The EEG baseline model reconstructs, cleans, and upscales data from all random channel structures. It builds on the ZUNA1, Zyphra’s pre-opened EEG base model.
The main change is flexibility, not a jump in raw precision. The actual EEG recording is messy. Sessions vary in length, and stations get loud or stop mid-session. The montages range from four-electrode headbands to 256-channel research caps. ZUNA1 only processed five-second fixed segments. ZUNA1.1 accepts inputs of variable length from 0.5 to 30 seconds.
What is ZUNA1.1?
To understand that flexibility, start with what the model does.
ZUNA1.1 is a 380M parameter latent autoencoder of scalp-EEG signals. Given a subset of channels, it refers to the existing EEG segments and channels. Rebuild the missing ones. It also predicts new signals given the physical links in the skin.
The parameter count is unchanged from ZUNA1. It runs on a consumer GPU and performs acceptably on a CPU in most tasks. Weights sit on Hugging Face; The inference and preprocessing code resides on GitHub. Enter with pip install zuna. Zyphra also hosts the free EEG Playground browser, and submits all of this for research use only.
How Architecture Works
That flexibility depends on tokenization.
ZUNA is a transformer-decoder autoencoder. It cuts each channel into 0.125 second segments, which is 32 samples at 256 Hz. Each part becomes a token with a constant value. Tokens are sorted by channel × time.
Spatial encoding is the main idea. Each token carries a 4D encoding that rotates over (x, y, z, t). That is the electrode’s 3D scalp coordinate and its coarse time reference. Because the location, not the array index, tells the model where the channel resides, ZUNA is channel-agnostic. It accepts any electrode configuration, and can generate signals in previously unrecorded areas. That capability allows the channel to be calibrated arbitrarily by location.
The encoder compresses the signal into a mask. That hidden sets the decoder with the standard adaptive-RMS. The decoder is trained with the purpose of a fixed flow. ZUNA1.1 structural changes target training stability, as standard additional layers.
What has changed in ZUNA1
As the architecture remained close, the difference came from the training.
1. Variable length input (0.5–30 seconds): ZUNA1.1 sample segment length per training example, averaged on a 0.125 s token grid. Height is plotted in four bins, from shortest to tallest. The average range of 1.5–10 s is an oversample, as it is the most common operating range. Because token counts vary, Zyphra packs multiple segments per batch into a fixed budget. Variable attention with a sample-aware mask sets tokens to every sample. So one model uses a 0.5 s clip and a 30 s stretch without reconditioning.
2. A rich mix of reconstruction activities: ZUNA1 trained on a single dropout pattern: completely random channels. ZUNA1.1 four trains. The first is to abandon the entire channel, including small montages and dead electrodes. The second is removing the short period that extends across all channels. The third removes those wires from certain channels only, covering gaps in space and time. The fourth one scatters the missing values at individual points.
3. Quality conscious pre-processing and large corpus: ZUNA1 made channel quality calls at the entire recording level, discarding a usable signal. ZUNA1.1 instead calculates a per-channel average, quality score per second, averaged over load time. That increased the corpus from about 2M to about 3.5M channel hours of public EEG data. The Zyphra team also precomputes two separate filters for each recording: a 0.1–45 Hz band, and a 0.01 Hz highpass and notch. Generalization across preprocessing techniques is a stated goal, not a limited effect.
Results
Therefore, the question is whether the cost variable is accurate.
For the captured functions, ZUNA1.1 achieves better or essentially the same NMSE for reconstruction as ZUNA1. Both clearly outperform the classical spherical-spline interpolation from MNE. For a fair comparison, those test sets used exactly five second samples.
Zyphra also conducts region-based testing. Electrodes from one area of the brain are removed, then reconstructed from the remaining seven. That setup makes a lot more sense than random channel drops. ZUNA1.1 outperforms both spherical-spline and ZUNA1 there.
Interactive Descriptor
To make those machines point, the demo below animates the pipeline at the end.
