Zyphra Releases ZUNA1.1: An Apache 2.0 EEG Foundation Model With Variable-Length Inputs From 0.5 To 30 Seconds

Zyphra just put out ZUNA1.1, the second version of its EEG foundation model, and it's free under Apache 2.0. If you've ever worked with real EEG data, you know why this matters: sessions run different lengths, electrodes fall off mid-recording, and setups range from a cheap four-electrode headband to a 256-channel research cap. ZUNA1, released earlier, could only chew through fixed five-second windows. ZUNA1.1 accepts anything from 0.5 to 30 seconds, which sounds like a small tweak but actually removes a real bottleneck for anyone trying to use this on messy, real-world recordings.

Under the hood, it's still a 380-million-parameter transformer encoder-decoder built as a diffusion autoencoder, so the parameter count hasn't budged from version one. It slices each channel into 32-sample, 0.125-second tokens and tags every token with a 4D rotary positional encoding based on the electrode's actual x, y, z scalp coordinates plus a time index. That's the trick that lets the model work with any electrode layout at all, and even generate plausible signal at scalp positions that were never recorded in the first place. The decoder runs on a rectified-flow objective, same as before.

The real story here is training, not architecture. Zyphra expanded the corruption tasks the model learns from four ways instead of one: dropping whole channels, wiping short time chunks across all channels, wiping chunks from just some channels, and scattering random missing points. They also junked the old whole-recording quality filter in favor of scoring quality per channel, per second, which let the training set balloon from about 2 million to roughly 3.5 million channel-hours of public EEG data. Two preprocessing variants — a 0.1 to 45 Hz bandpass and a 0.01 Hz highpass with notch filtering — get precomputed for every recording now, too.

Does any of this cost accuracy for the sake of flexibility? Apparently not much. On matched five-second evaluation windows, ZUNA1.1 matches or beats ZUNA1's reconstruction error, and both stomp the classical spherical-spline interpolation method from MNE. More telling is the region-deletion test, where Zyphra wipes out every electrode in one brain region and asks the model to rebuild it from the remaining seven regions — a much harder, more realistic failure mode than randomly dropping channels. ZUNA1.1 wins there too.

Zyphra also cleaned up the tooling. There's now a one-line reconstruct_fif function that works straight on .fif files, lets you name specific bad channels or time spans to repair, and can upsample a sparse headband recording by requesting additional standard-position electrodes that were never actually measured. It's research-use-only for now, but with weights on Hugging Face and code on GitHub, it's a genuinely usable release rather than a paper with a locked-up demo.