Most language models learn by guessing the next word, millions of times over, until meaning emerges from sheer scale. A small Indian team took the opposite route: start with meaning itself. NeuroDX, the deep-tech startup now becoming Manas AI, has built Sphota, a 240-million-parameter language model that treats the event behind the words as the thing to be modelled — and the company says it already outperforms a comparable model from Liquid AI while using 50 times less training data.
Illustration of brain-signal AI: glowing EEG traces around a brain model. (Illustration)
From brainwaves to language
Sphota did not begin as a language model at all. NeuroDX began in healthcare, building AI that reads electroencephalogram (EEG) signals — the electrical activity recorded from the scalp. Brain signals carry rich information about what the brain is doing, but they are far harder to interpret than a photograph or an MRI scan, and they do not break neatly into words.
That was the spark. The company says conventional models divide text into tokens and learn by predicting what comes next, but EEG signals have no natural token boundaries — splitting them requires arbitrary decisions that limit how well a model can understand the underlying information. While building a foundation model for the brain with support from the Ministry of Electronics and Information Technology (MeitY) and the IndiaAI Mission, founder and CEO Siddharth Panwar and his team asked whether they could move beyond tokens entirely and model meaning directly.
The answer became Sphota, previewed at Cypher 2026 in Bengaluru on 9 October. The name comes from the linguistic concept associated with Bhartrihari, the Indian philosopher and grammarian who argued that meaning is grasped as a whole, not assembled word by word. Two sentences can describe the same event in different words — someone gives another person a book, or the recipient receives a book from the giver — and Sphota is designed to recognise the shared event from the outset, rather than learning the equivalence only after enormous amounts of training.
More Read
Small model, striking numbers
The company says Sphota scored 52.4% against Liquid AI’s 52.1% across an average of about 11 benchmarks, and that it reached that performance with 50 times less training data. The company also says the architecture delivers 4.5 times better memory efficiency, letting the model retain substantially more conversational context while using less memory — a difference that matters as conversations grow longer and context windows grow larger.
Panwar, an assistant professor at IIT Mandi, frames the work as a different kind of ambition for Indian AI. Rather than chasing the same transformer-based path as the world’s largest labs, the company says, a small team can leap ahead by building something orthogonal — an architecture grounded in how the brain itself handles meaning. The company has been working on Sphota for only a few months with a small team, and the company says a larger model of between 1 billion and 2 billion parameters is targeted for release around December, potentially accessible through a chatbot on the company’s website.
The brain model behind it
Sphota’s roots run deep into NeuroDX’s healthcare work. MANAS-1, the company’s 400-million-parameter Brain Language Foundation Model, was trained on 60,000 hours of brainwave recordings from more than 25,000 patients and released as open source on Hugging Face. Developed by Intellihealth (NeuroDx) under neurologist Dr Puneet Agarwal, a former professor at the All India Institute of Medical Sciences, MANAS-1 is designed to understand what the company calls the basic language of the brain — a foundation on which disease-specific tools can be built.
The company says MANAS-1 has achieved more than 95% accuracy in identifying biomarkers related to epilepsy and other mental-health disorders, and the model has been submitted to the Government of India for evaluation under emerging standards for sovereign AI assets. Its successor, MANAS-2, is described on the company’s website as a world-class EEG foundation model, powering a suite of healthcare solutions including EpiDx and MindDx.
The healthcare goal is strikingly practical: EEG-based screening that flags people who may be at risk of neurological conditions including dementia, epilepsy and depression, using a small, relatively inexpensive device that can be operated without highly specialised personnel. The company says its AI models generate a risk profile to help an MBBS doctor decide whether to refer a patient to a specialist — screening designed to reach primary health centres and villages where neurological specialists are scarce. The company says the models are undergoing evaluation and could be ready for deployment by the end of 2026 or early 2027.
Building India’s own frontier
NeuroDX is among the companies selected to develop sovereign AI models under the IndiaAI Mission. The company says it was allocated 368 H200 GPUs for 18 months and ₹12.5 crore to develop a 20-billion-parameter multimodal EEG model — backing that underscores both the opportunity and the constraints facing Indian teams building frontier technology with far less compute than global giants.
Panwar’s answer to that constraint is to refuse it as an excuse. As the company rebrands itself as Manas AI and moves from a healthcare-focused business toward a foundation-model company, its ambition is to build general-purpose intelligence from India — and to show Indian engineers that frontier technology is something they can create, not only work for elsewhere. The company says its brain foundation models are publicly available, while models developed with MeitY and IndiaAI Mission support are intended to be open source, even as the Sphota architecture itself stays proprietary for now.
That is the heart of NeuroDX’s story: a team that learned to read the brain discovered a new way to build AI — and turned it into a model that does more with less. Meaning first, words second, and a 240-million-parameter proof that India can chart its own path to the frontier.
Keep reading: The Carbon You Can’t See: Inside Fitsol, the Gurugram Startup Turning India’s Supply Chains Green · Quantum Keys on a Beam of Light: Inside NavTech’s OpticSpectra 10G QKD · More Industrial Automation & Technology stories
Like stories like this? Add Exceptional Technology as a preferred source on Google to see more of our reporting in your search results and Discover feed.
