MANAS-1: India's First Brain-Wave Foundation Model

India has built a foundation model for the brain. NeuroDX (Intellihealth) unveiled MANAS-1 at the India AI Impact Summit around 23 February 2026: a 400-million-parameter foundation model trained on 60,000 hours of EEG recordings from more than 25,000 patients, released open-source on Hugging Face. The company reports accuracy above 95% on epilepsy-related biomarkers in research settings, and MANAS-1 received compute support under the IndiaAI Mission. Notably, it was selected as one of 12 IndiaAI Mission "Champions", alongside text-model builders such as Sarvam and BharatGen, and is the only physiological foundation model on that list. It was built by founders Dr. Siddharth Panwar (IIT Delhi and Stanford) and Kailash Sati, with former AIIMS neurologist Dr. Puneet Agarwal as a clinical collaborator, and the company says it remains bootstrapped.

Foundation models meet the brain

The conceptual move is the same one that produced large language models, redirected at neural data. A language model learns the statistical structure of language by training on enormous quantities of text; MANAS learns the structure of brain activity by training on enormous quantities of EEG, the electrical signal recorded from electrodes on the scalp. The payoff is the same too. Instead of building a separate, narrowly trained model for each task, one for epilepsy detection, another for sleep staging, a third for a particular cognitive marker, a single pretrained base can be fine-tuned for many downstream applications using far less task-specific data than training from scratch would need.

That general-purpose quality is what makes a foundation model more than the sum of its training tasks. The model learns reusable representations of what brain activity looks like, and those representations transfer to problems it was never explicitly trained on. Applied to the brain, that is a genuinely new capability rather than an incremental one.

Why EEG, and why this is hard

EEG is an unusually good target for this approach, for a practical reason: it is cheap and widely available but genuinely hard to interpret. An EEG machine costs a fraction of an MRI scanner and is found in hospitals across the country, yet reading the squiggling traces it produces well enough to spot the signatures of epilepsy or other disorders requires scarce specialist expertise that most clinics simply do not have. A capable foundation model that can surface those biomarkers from raw traces could extend that expertise to wherever the recording hardware exists, which in India is almost everywhere the specialists are not.

The difficulty is that EEG is messy. The signal is faint, contaminated by muscle movement and electrical noise, and varies enormously between individuals and even within one person over time. Assembling 60,000 hours of it from more than 25,000 patients, and training a model large enough to learn the underlying patterns rather than the noise, is a serious undertaking, and the scale is what makes the result notable.

Sovereign, physiological, and open

MANAS-1 is most interesting for what it represents within India's AI strategy. The country's sovereign-AI push has so far been dominated by language: open-weight large language models trained for Indian languages and contexts, which is what Sarvam, BharatGen and their peers are building. MANAS extends that bet from language into physiology and health. It is built on Indian patient data, trained with public compute through the IndiaAI Mission, and released with open weights so others can build on it. Being the only physiological model among the mission's official champions makes it a deliberate signal that "sovereign AI" should include the body and the clinic, not only text and chat.

That framing has real consequences. A physiological foundation model owned and openly published in India is infrastructure others can fine-tune for Indian healthcare, rather than a black box licensed from abroad, and it keeps both the data and the capability inside the country.

The caveats

The caveats deserve emphasis, because medical AI invites overclaiming and the consequences of believing the hype are measured in misdiagnoses. The above-95% accuracy figure is a research-setting result reported through health-technology outlets, not a regulatory clearance or a peer-reviewed clinical endpoint, and the path from a strong benchmark to a validated, approved clinical tool is long, expensive and heavily regulated. EEG models also inherit the biases of their training data, so the composition of those 60,000 hours, which conditions, which demographics, which recording equipment, will shape where the model works well and where it fails. The company has promised a far larger MANAS-2 running into the billions of parameters, which is a roadmap, not a result.

For now, the significance is strategic rather than clinical. India has planted a flag in physiological foundation models, an area most of the world's sovereign-AI programmes have not yet touched, and it has done so openly and on home-grown data. Whether MANAS becomes a tool clinicians actually rely on will depend on the unglamorous years of validation that come next.