{"id":105189,"date":"2026-07-18T19:30:09","date_gmt":"2026-07-18T19:30:09","guid":{"rendered":"https:\/\/youzum.net\/zyphra-releases-zuna1-1-an-apache-2-0-eeg-foundation-model-with-variable-length-inputs-from-0-5-to-30-seconds\/"},"modified":"2026-07-18T19:30:09","modified_gmt":"2026-07-18T19:30:09","slug":"zyphra-releases-zuna1-1-an-apache-2-0-eeg-foundation-model-with-variable-length-inputs-from-0-5-to-30-seconds","status":"publish","type":"post","link":"https:\/\/youzum.net\/it\/zyphra-releases-zuna1-1-an-apache-2-0-eeg-foundation-model-with-variable-length-inputs-from-0-5-to-30-seconds\/","title":{"rendered":"Zyphra Releases ZUNA1.1: An Apache 2.0 EEG Foundation Model With Variable-Length Inputs From 0.5 To 30 Seconds"},"content":{"rendered":"

This week, Zyphra released ZUNA1.1 under the Apache 2.0 license<\/a><\/strong>. The EEG foundation model reconstructs, denoises, and upsamples data across arbitrary channel layouts. It builds on ZUNA1, the Zyphra\u2019s earlier open EEG foundation model.<\/p>\n

The main change is flexibility, not a jump in raw accuracy. Real EEG recordings are messy. Sessions vary in length, and channels go noisy or drop out mid-session. Montages range from four-electrode headbands to 256-channel research caps. ZUNA1 processed only fixed five-second segments. ZUNA1.1 accepts variable-length inputs from 0.5 to 30 seconds.<\/p>\n

What is ZUNA1.1?<\/strong><\/h2>\n

To understand that flexibility, start with what the model does.<\/p>\n

ZUNA1.1 is a 380M-parameter masked diffusion autoencoder for scalp-EEG signals. Given a subset of channels, it denoises existing EEG segments and channels. It reconstructs missing ones. It also predicts novel channel signals given physical coordinates on the scalp.<\/p>\n

The parameter count is unchanged from ZUNA1. It runs on a consumer GPU and works acceptably on CPU for many workloads. Weights sit on Hugging Face; inference and preprocessing code sit on GitHub. Install with pip install zuna<\/code>. Zyphra also hosts a free browser EEG Playground, and ships all of this for research use only.<\/p>\n

How The Architecture Works<\/strong><\/h2>\n

That flexibility rests on tokenization.<\/p>\n

ZUNA is a transformer encoder\u2013decoder diffusion autoencoder. It slices each channel into 0.125 second segments, which is 32 samples at 256 Hz. Each segment becomes a continuous-valued token. Tokens are serialized in channel \u00d7 time order.<\/p>\n

The positional encoding is the key idea. Each token carries a 4D rotary positional encoding over (x, y, z, t). That is the electrode\u2019s 3D scalp coordinate along with its coarse-time index. Because position, not array index, tells the model where a channel sits, ZUNA is channel-agnostic. It accepts any electrode layout, and can generate signals at positions never recorded. That capability enables arbitrary channel upsampling by location.<\/p>\n

The encoder compresses the signal into a latent. That latent conditions the decoder via adaptive-RMS norm. The decoder is trained with a rectified-flow objective. ZUNA1.1\u2019s architectural changes targeted training stability, such as added normalization layers.<\/p>\n

What Changed From ZUNA1<\/strong><\/h2>\n

Since the architecture stayed close, the differences come from training.<\/p>\n

1. Variable-length inputs (0.5\u201330 seconds)<\/strong>: ZUNA1.1 samples a segment length per training example, snapped to the 0.125 s token grid. Lengths are drawn across four bins, from very short to long. The middle 1.5\u201310 s range is oversampled, since it is the most common operating point. Because token counts vary, Zyphra packs multiple segments per batch up to a fixed budget. Flex attention with a sample-aware mask stops tokens attending across samples. One model therefore serves a 0.5 s snippet and a 30 s stretch without reconfiguration.<\/p>\n

2. A richer mixture of reconstruction tasks<\/strong>: ZUNA1 trained on one dropout pattern: uniformly random whole channels. ZUNA1.1 trains on four. The first is whole-channel dropout, covering sparse montages and dead electrodes. The second removes short time stretches across every channel. The third removes those stretches from only some channels, clustering gaps in space and time. The fourth scatters missing values across individual points.<\/p>\n

3. Quality-aware preprocessing and a bigger corpus<\/strong>: ZUNA1 made channel-quality calls at the whole-recording level, discarding usable signal. ZUNA1.1 instead computes a per-channel, per-second quality score, thresholded at load time. That grew the corpus from roughly 2M to roughly 3.5M channel-hours of public EEG data. Zyphra team also precomputes two filter variants per recording: a 0.1\u201345 Hz bandpass, and a 0.01 Hz highpass along with notch. Generalizing across preprocessing strategies is a stated goal, not a benchmarked result.<\/p>\n

The Results<\/strong><\/h2>\n

Consequently, the question is whether flexibility cost accuracy.<\/p>\n

On held-out tasks, ZUNA1.1 reaches better or essentially the same reconstruction NMSE as ZUNA1. Both clearly outperform classical spherical-spline interpolation from MNE. For fair comparison, those evaluation sets used exactly five-second samples.<\/p>\n

Zyphra also ran a region-based test. Electrodes from one brain region are deleted, then reconstructed from the remaining seven. That setup is more realistic than random channel dropping. ZUNA1.1 outperforms both spherical-spline and ZUNA1 there.<\/p>\n

Interactive Explainer<\/strong><\/h2>\n

To make those mechanics concrete, the demo below animates the pipeline end to end.<\/p>\n

\n

ZUNA1 vs ZUNA1.1<\/strong><\/h2>\n<\/p>

Taken together, the releases differ mostly in training, not architecture.<\/p>\n

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Attribute<\/th>\nZUNA1<\/th>\nZUNA1.1<\/th>\n<\/tr>\n<\/thead>\n
Parameters<\/td>\n380M<\/td>\n380M<\/td>\n<\/tr>\n
Architecture<\/td>\nTransformer encoder\u2013decoder diffusion autoencoder<\/td>\nSame, plus extra normalization layers<\/td>\n<\/tr>\n
Input length<\/td>\nFixed 5 s<\/td>\n0.5\u201330 s, snapped to 0.125 s grid<\/td>\n<\/tr>\n
Token<\/td>\n0.125 s \/ 32 samples at 256 Hz<\/td>\nSame<\/td>\n<\/tr>\n
Positional encoding<\/td>\n4D RoPE over (x, y, z, t)<\/td>\nSame<\/td>\n<\/tr>\n
Decoder objective<\/td>\nRectified flow<\/td>\nRectified flow<\/td>\n<\/tr>\n
Dropout schemes in training<\/td>\n1 (uniform random whole-channel)<\/td>\n4 (channel, time, channel\u00d7time, scattered)<\/td>\n<\/tr>\n
Training corpus<\/td>\n~2M channel-hours<\/td>\n~3.5M channel-hours<\/td>\n<\/tr>\n
Quality filtering<\/td>\nWhole-recording level<\/td>\nPer-channel, per-second score at load time<\/td>\n<\/tr>\n
Preprocessing variants<\/td>\nSingle<\/td>\nTwo (0.1\u201345 Hz bandpass; 0.01 Hz highpass + notch)<\/td>\n<\/tr>\n
License<\/td>\nApache 2.0<\/td>\nApache 2.0<\/td>\n<\/tr>\n
Reconstruction NMSE<\/td>\nBaseline<\/td>\nEqual or better<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

Running It<\/strong><\/h2>\n

Turning to practice, reconstruct_fif<\/code> runs directly on .fif<\/code> files with no .pt<\/code> round-trip. The older four-step pipeline still ships alongside it.<\/p>\n

\n
\n
\n
\n
<\/div>\n
<\/div>\n
<\/div>\n<\/div>\n
Copy Code<\/span>Copied<\/span>Use a different Browser<\/span><\/a><\/div>\n<\/div>\n
from zuna import reconstruct_fif\n\nreconstruct_fif(\n    input_dir=\"fif_in\",\n    output_dir=\"fif_out\",\n    figures_dir=\"figures\",\n    gpu_device=0,                            # GPU id, or \"\" for CPU\n    segment_sec=5.0,                         # window length; default is 5.0, not the full 30 s\n    montage=\"standard_1020\",                 # fallback, used only if the file has no positions\n    repair_channels=[\"Cz\"],                  # channel(s) to fully reconstruct\n    target_channel_count=[\"Fz\", \"Pz\"],       # add\/upsample new channels by name (or an int for auto)\n    bad_segments=[(5, 6), (10, 11, \"C3\")],   # mark time spans bad (all channels, or one)\n    sample_steps=50,                         # diffusion steps; note: not \"diffusion_sample_steps\"\n)<\/code><\/pre>\n<\/div>\n<\/div>\n
Note the defaults. segment_sec<\/code> is 5.0, so the 0.5\u201330 s range needs setting explicitly. Electrode positions are read from the file itself. The montage<\/code> argument is only a fallback when positions are absent, and channels without 3D coordinates are dropped.<\/p>\n
The reconstruction target is a union. It combines the file\u2019s own MNE bad channels and BAD_<\/code> annotations with anything requested above. Two directories are written. full_reconstruction\/<\/code> holds model output everywhere. hybrid\/<\/code> keeps the original and infills only inferred cells, plus a _mask.npz<\/code>.<\/p>\n
Use Cases With Examples<\/strong><\/h2>\n
Because masking is now flexible, several practical patterns open up.<\/p>\n