India’s frontier AI ambitions may have found an unlikely contender in a 240-million-parameter language model that challenges conventional approaches to building AI.
NeuroDX, an Indian startup backed by the IndiaAI Mission that grew out of brain research, says its new model, Sphota, has outperformed a comparable model from Liquid AI across an average of about 11 benchmarks, while requiring significantly less training data.
Sphota scored 52.4% against Liquid AI, which recorded 52.1%, according to the company’s reported benchmark results. Notably, NeuroDX CEO Siddharth Panwar claims Sphota achieves comparable performance with 50 times less training data and offers 4.5 times better memory efficiency.
The results are preliminary and have not yet been established through independent public evaluations. However, they point to a potentially different approach to building language models, one that attempts to capture meaning directly rather than relying primarily on predicting the next word.
“We didn’t go out to build an efficient small model. We just had an architecture, and we started with small sizes because that’s the size we can play at,” Panwar said in an exclusive interview with AIM’s Front Page.
A New Architecture Built Around Meaning
Sphota’s origins lie in an effort to understand the human brain.
NeuroDX initially focused on healthcare, developing AI models that interpret electroencephalogram (EEG) signals recorded from the scalp. These signals contain information about brain activity but are difficult to interpret directly, unlike medical images such as MRI scans.
While developing a foundation model for the brain with support from the Ministry of Electronics and Information Technology (MeitY) and the IndiaAI Mission, Panwar and his team encountered a problem with conventional AI architectures.
Most language models divide text into tokens and learn patterns by predicting what comes next. However, EEG signals do not have the same natural divisions as written language. Splitting them into tokens requires arbitrary decisions, which can limit how effectively a model interprets the underlying information.
The team began exploring whether it could move beyond tokens and model the meaning embedded in information instead.
The resulting architectural innovations proved applicable beyond brain signals, leading the company to experiment with language models.
Sphota takes its name from the linguistic concept associated with Bhartrihari, an Indian philosopher and linguist who argued that meaning is apprehended as a whole rather than assembled merely by processing individual words.
Panwar believes conventional language models can eventually arrive at an understanding of meaning, but often require enormous amounts of training data and computing power to do so.
Consider two sentences describing the same event: someone gives another person a book, and the recipient receives a book from the giver. Although the wording differs, the underlying event remains unchanged.
A conventional model learns to recognise such relationships through extensive training. Sphota, by contrast, makes identifying these underlying similarities a design principle from the outset.
“What you’re looking to build is that main layer not only about the tokens but consciously as an engineering architecture level you are trying to see how can you build this later,” Panwar said.
The company argues that this approach could help smaller models develop certain capabilities with less training data. Whether those gains persist across a wider range of tasks and larger models remains to be established.
50x Less Training Data, 4.5x Better Memory Efficiency
The most consequential claims surrounding Sphota concern the cost of training and running AI models.
According to Panwar, the 50x figure compares the amount of training data used by Sphota and Liquid AI’s model to achieve similar benchmark performance. The 4.5x memory-efficiency claim relates to how much conversational context a model can retain under comparable conditions.
Longer conversations and larger context windows require models to manage increasing amounts of information, placing additional demands on memory resources.
Panwar said Sphota’s architecture could allow it to retain substantially more context than a comparable model while using less memory.
However, these figures represent the company’s reported comparisons, not independently verified results across all models and workloads. The eventual gains will depend on how the architecture performs as it scales and how it compares with other systems under equivalent testing conditions.
Panwar also cautioned that the efficiency gains observed so far may not translate directly into the same numbers when the model becomes widely available.
The company has been working on Sphota for only a few months with a small team. It is now developing a larger model with between 1 billion and 2 billion parameters, with Panwar targeting a release around December. He said users could potentially access it through a chatbot on the company’s website.
The next model will be an important test of whether Sphota’s early results hold up at larger scales.
NeuroDX Becomes MANAS AI
Alongside its model ambitions, NeuroDX is preparing to change its identity.
Panwar revealed the company is rebranding itself as Manas AI, reflecting its transition from a healthcare-focused business towards a foundation model company.
The startup remains involved in neurological healthcare, but its broader ambition is to develop general-purpose intelligence and build a global technology company from India.
Panwar said the company’s brain foundation models are publicly available, while models developed with support from MeitY and the IndiaAI Mission are intended to be open source.
However, the company plans to keep the underlying Sphota architecture proprietary for now, allowing it to build a competitive advantage and attract the resources needed to develop the technology further.
“We are just making sure that this architecture stays proprietary,” Panwar said.
Why India Cannot Afford to Abandon Frontier AI
Sphota’s emergence also raises a larger question: Can Indian companies compete in frontier AI without access to the computing resources and capital available to global technology giants?
NeuroDX is among the companies selected to develop sovereign AI models under the IndiaAI Mission. Panwar said the company was allocated 368 H200 GPUs for 18 months and ₹12.5 crore to develop a 20-billion-parameter multimodal EEG model.
The allocation underscores the constraints facing Indian AI developers attempting to build ambitious models with limited resources.
Panwar acknowledged the frustration of having engineers ready to experiment but insufficient GPUs to pursue every research direction. However, he argued that the shortage should not force India to abandon the frontier AI race.
“Giving up on the frontier just because we don’t have enough risk capital is suicidal,” he said.
For Panwar, the case for building frontier models extends beyond commercial returns. Developing these systems domestically would allow Indian researchers to understand how advanced AI works, investigate unexpected model behaviour and develop appropriate safeguards rather than relying entirely on technology built elsewhere.
Established AI laboratories have invested heavily in transformer-based systems, giving startups an opportunity to investigate approaches that do not depend on the same underlying assumptions.
“You can’t leaprog by chasing them. You can leaprog them by building something completely orthogonal in kind of the technological directions they were taking and then find yourself up ahead of them,” he said.
Panwar said the company is looking to raise more than $10 million, although the interview did not establish a completed funding round.
From Brain Signals to Village-Level Healthcare
Despite its new focus on foundation models, Manas AI’s healthcare ambitions remain central to its work.
The company is developing EEG-based screening systems designed to identify people who may be at risk of neurological conditions, including dementia, epilepsy and depression.
The system records brain signals using a small, relatively inexpensive device that can be operated without highly specialised personnel. Its AI models then generate a risk profile to help an MBBS doctor decide whether to refer a patient to a specialist.
The aim is to extend screening to primary healthcare centres and villages where access to neurological specialists is limited.
Panwar said the models were undergoing evaluation and could be ready for deployment by the end of 2026 or early 2027. Wider deployment would depend on working with the government and other healthcare stakeholders.
The company has collected carefully labelled EEG data over an extended period, with the required ethical approvals and consent, to develop and fine-tune these models.
Sphota’s immediate test is whether its reported benchmark and efficiency gains survive independent scrutiny and scale beyond 240 million parameters. Its longer-term test is whether a small Indian team can turn an architectural breakthrough into a competitive foundation model.
For Panwar, the ambition goes beyond building another AI startup. He wants Indian engineers to believe they can create frontier technology rather than merely work for companies that already have.
“The whole purpose of Manas is to show that there is absolutely nothing that we can’t do,” he said.
