ZUNA EEG Foundation Model

Source: ZUNA on zyphra.com (02/18/2026) — Chris Warner, Jonas Mago, Jonathan Huml, Beren Millidge. Paper: ZUNA: Flexible EEG Superresolution with Position-Aware Diffusion Autoencoders (arXiv:2602.18478).

The ZUNA article is not about a model that "reads thoughts into text". It is a foundation model for scalp EEG signals: denoising, filling in dropped channels, and "super-resolving" a sparse montage into a denser one. Thought-to-text is Zyphra's long-term direction; ZUNA is only the signal-processing foundation layer.

What the Article Says

Non-invasive EEG is cheap and widely available, but real-world data is messy: channels drop out, motion artifacts corrupt signals, and montages vary wildly (a 4-channel Muse headband vs a 256-channel research cap). The standard tool in MNE is spherical spline interpolation — geometric interpolation on a sphere that learns no prior from real data. It works when few channels are missing but breaks down past ~75% dropout.

ZUNA replaces geometric interpolation with a model trained on ~2 million channel-hours of public EEG (208 datasets), so it can:

• denoise the channels that exist
• reconstruct channels that were lost
• predict the signal at electrode positions never recorded, given only their 3D scalp coordinates

Practical goal: rescue "unusable" recordings, bring cheap headsets closer to lab quality, and stop depending on a fixed montage. The marketing framing — multimodal AI after text/audio/vision is thought-to-text via non-invasive BCI in service of "human-aligned superintelligence" — is a mission statement, not a demonstrated result.

Model Facts

GitHub now defaults to ZUNA1.1. Weights: Zyphra/ZUNA (original), Zyphra/ZUNA1.1. Code: github.com/Zyphra/ZUNA. Install: pip install zuna. Runs on consumer GPUs or CPU — GitHub claims <1 GB VRAM. A playground on Zyphra Cloud accepts .fif uploads.

Not an LLM. It does not chat and does not decode inner language. Input and output are EEG time series.

Architecture

Diffusion autoencoder + transformer encoder–decoder.

1. Each channel is cut into 0.125 s chunks (32 samples at 256 Hz) → continuous tokens (not discrete GPT-style tokens).
2. Tokens form a 1D sequence over channel × time.
3. Positional encoding is 4D RoPE over (x, y, z, t): electrode coordinates on the scalp plus raw time. The model doesn't memorize "channel 7" — it learns "this physical location at this moment". That is what makes it montage-agnostic and able to upsample new positions.
4. The encoder compresses into a latent space; the decoder diffuses/reconstructs the masked parts. Trained with masked reconstruction plus very heavy channel dropout.

ZUNA1.1 adds Sandwich-norm, QK-norm, a 3-stage training curriculum (~580k steps), and a mix of 4 corruption types to better match real-world EEG.

Benchmarks (as reported)

Against MNE's default spherical spline, on validation sets plus unseen datasets / different preprocessing pipelines:

• ZUNA wins clearly, and the gap widens as dropout increases.
• With >75% of channels missing: ZUNA beats spline on every dataset they report.
• ZUNA1.1: NMSE equal or better than ZUNA1 on 5 s windows while handling more corruption types. NMSE ~0.5 when masking 15–20% of tokens — described as "looks close to ground truth".

Caveat: this is waveform reconstruction, not decoding accuracy for motor imagery, speech, or thought. The paper emphasizes better cross-dataset generalization than earlier EEG deep-learning models tied to a single montage.

How to Read "Thought-to-Text"

The pipeline they imagine:

raw EEG, sparse, noisy
    → ZUNA (clean / infill / upsample)
        → a future decoding model (not ZUNA)
            → text / agent

ZUNA only completes stage 1. The semantic decoding stage does not exist in this work. Their own disclaimer: research only, not for clinical use, and reconstructed signals can be "hallucinated" (plausible but not ground truth).

Usage

from zuna import reconstruct_fif

reconstruct_fif(
    input_dir="fif_in",
    output_dir="fif_out",
    figures_dir="figures",
    gpu_device=0,                 # "" for CPU
    repair_channels=["Cz"],
    target_channel_count=["Fz", "Pz"],
    bad_segments=[(5, 6), (10, 11, "C3")],
)

Expected input: MNE .fif files, 256 Hz sample rate, with electrode coordinates (or a standard_1020 montage assigned).

References

• Article: zyphra.com/our-work/zuna
• ZUNA paper: arXiv:2602.18478
• ZUNA1.1 paper: arXiv:2607.27308

In one sentence: ZUNA is a 380M open-source model that repairs and "upscales the resolution" of EEG using electrode coordinates — a foundation for non-invasive BCI, not a mind-reading machine.

See also: Transformer Architecture · Attention Mechanism · Code World Model