Zyphra has taken a significant step in brain signal processing with the release of ZUNA1.1, an open-source EEG foundation model under Apache 2.0 license. This model, improving on its predecessor ZUNA1, introduces unprecedented flexibility by accepting input segments from 0.5 to 30 seconds, compared to the fixed 5-second segments of the previous version. With 380 million parameters, ZUNA1.1 is a masked diffusion autoencoder that reconstructs, denoises, and upsamples electroencephalogram signals, even on channels that were never recorded. Its ability to work with any electrode layout, thanks to a 4D rotary positional encoder over (x, y, z, t) coordinates, makes it a universal tool for researchers and businesses.
The architecture of ZUNA1.1 is based on a transformer encoder–decoder with a rectified flow objective. Each channel is divided into continuous tokens of 0.125 seconds (32 samples at 256 Hz), serialized in channel × time order. The key innovation is the positional encoding: each token carries its 3D scalp coordinate along with a coarse time index, allowing the model to be channel-agnostic. This means it can generate signals at locations never measured before, enabling arbitrary channel upsampling. Additionally, the model includes extra normalization layers to improve training stability.
The improvements over ZUNA1 are substantial. Training now uses variable-length segments, sampled across four bins favoring the 1.5 to 10 second range, the most common in practice. Three new dropout patterns were added: removal of time spans on all channels, removal of spans on some channels only, and scattered missing values on individual points. The training corpus grew from 2 to 3.5 million channel-hours, thanks to per-channel, per-second quality filtering instead of whole-recording level. Two preprocessing variants are also included: a 0.1–45 Hz bandpass filter and a 0.01 Hz highpass with notch. All this without sacrificing accuracy: in tests with 5-second samples, ZUNA1.1 matches or exceeds ZUNA1, and clearly outperforms classical spherical-spline interpolation from MNE.
From a business perspective, this breakthrough opens doors to applications in healthcare, brain-computer interfaces, and neuroscience. However, deploying AI models like ZUNA1.1 in production environments requires robust and scalable software architecture. This is where Q2BSTUDIO brings its expertise in custom software development. Integrating an EEG model into a clinical system, for example, involves creating data pipelines, managing sensitive information security, and deploying cloud infrastructure. Our team can design cloud AWS/Azure solutions that process signals in real time, combining the model with AI to detect anomalies or predict cognitive states.
Cybersecurity is another critical pillar. Biomedical data is highly regulated, and any solution handling EEG must comply with standards like GDPR or HIPAA. Q2BSTUDIO offers cybersecurity services to ensure that the infrastructure behind these models is protected against unauthorized access and data leaks. Furthermore, integration with BI/Power BI systems allows brain patterns to be visualized on interactive dashboards, facilitating decision-making in research or diagnosis.
The flexibility of ZUNA1.1 is also key for automation of workflows in laboratories. For example, a system can receive recordings from different montages and automatically generate clean, upsampled signals without manual intervention. Combined with AI agents, intelligent assistants can suggest corrections or alert about artifacts. At Q2BSTUDIO, we work on automation solutions that integrate AI models with enterprise tools, reducing costs and accelerating research.
The model is distributed with weights on Hugging Face, inference code on GitHub, and installable via 'pip install zuna'. Zyphra also offers a free browser EEG Playground. For commercial use, scalability and customization requirements must be evaluated. Companies wishing to leverage ZUNA1.1 for their own products can benefit from Q2BSTUDIO's consulting in AI agents and custom software development. Whether for a neurotechnology startup or a hospital aiming to improve diagnostics, the combination of open foundation models and solid engineering is the winning formula.
In conclusion, ZUNA1.1 represents a milestone in democratizing EEG analysis. Being open source and flexible, it allows developers and scientists to innovate without technical limitations. However, true transformation comes when these capabilities are integrated into secure, scalable enterprise systems. Q2BSTUDIO is ready to accompany that journey, offering services from cloud deployment to custom user interfaces. The era of artificial intelligence applied to the human brain has just taken a new leap, and companies that ride this wave will lead the future of neurotechnology.
