India’s Frontier AI Race Gets a 240M Model That Beats Liquid AI

India’s artificial intelligence industry is gaining an unexpected contender in the frontier AI race. NeuroDX, an Indian startup backed by the IndiaAI Mission, has developed Sphota, a 240-million-parameter language model that it claims outperforms a comparable model from US-based Liquid AI across approximately 11 benchmarks.

According to results shared by the company, Sphota achieved an average benchmark score of 52.4%, compared with 52.1% for Liquid AI’s model. More significantly, NeuroDX claims its model achieved comparable performance using 50 times less training data and delivered 4.5 times better memory efficiency. These results are preliminary and have not yet been independently verified, but they highlight an alternative approach to building smaller and more efficient AI models.

What Is Sphota, India’s 240-Million-Parameter AI Model?

Sphota is a language model developed by NeuroDX, an Indian AI startup that initially focused on building artificial intelligence systems for analysing brain signals.

Unlike conventional large language models that primarily rely on predicting the next token, Sphota is built around an architectural approach that attempts to capture the underlying meaning of information.

The model contains 240 million parameters, making it considerably smaller than many modern foundation models developed by global AI companies.

Despite its relatively small size, NeuroDX claims Sphota has achieved competitive benchmark results against a comparable model from Liquid AI.

Here is a comparison of the reported results:

Metric	NeuroDX Sphota	Liquid AI Model
Parameters	240 million	Comparable model
Average benchmark score	52.4%	52.1%
Benchmarks evaluated	Approximately 11	Approximately 11
Training data requirement	Claimed 50x lower	Reference comparison
Memory efficiency	Claimed 4.5x better	Reference comparison
Independent verification	Not yet available	Not yet established

The figures are based on NeuroDX’s reported comparisons rather than independent evaluations. The company has not disclosed enough information to establish that the models were tested under fully equivalent conditions across every benchmark.

Nevertheless, the results have attracted attention because they suggest that smaller models may achieve competitive capabilities without requiring the enormous training datasets and computing resources typically associated with advanced AI development.

How Sphota Uses a Different Approach to AI Architecture

One of the most interesting aspects of Sphota is its underlying approach to language understanding.

Most modern large language models use transformer-based architectures. They divide text into smaller units called tokens and learn patterns by predicting what comes next.

For example, when a model receives the sentence “The boy is reading a book,” it learns to predict the next token based on the surrounding context.

As training continues across billions or trillions of tokens, the model develops increasingly sophisticated representations of language, relationships and reasoning patterns.

NeuroDX is exploring a different approach.

Instead of treating next-token prediction as the primary design principle, the company is attempting to build an architecture that captures meaning more directly.

Understanding Meaning Instead of Just Predicting Words

Consider two sentences:

• A person gives another person a book.
• Someone receives a book from another person.

Although the sentences use different words and grammatical structures, they describe essentially the same event.

A conventional language model learns these relationships through extensive training on large amounts of text.

Sphota aims to make identifying these underlying similarities a more fundamental part of its architecture.

The model takes its name from the linguistic concept of Sphota, associated with the ancient Indian philosopher and grammarian Bhartrihari.

The concept describes how meaning can be understood as a complete expression rather than simply being assembled from individual words.

NeuroDX CEO Siddharth Panwar believes this principle can help AI systems develop meaningful representations of information with fewer training resources.

The company is not claiming that conventional language models cannot understand meaning. Its argument is that explicitly designing for meaning could make the learning process more efficient.

From Brain Research to Language Models

Sphota’s development has an unusual origin.

NeuroDX initially focused on healthcare applications, particularly artificial intelligence systems capable of interpreting electroencephalogram (EEG) signals.

EEG technology records electrical activity in the brain through sensors placed on the scalp. These signals can help medical professionals investigate neurological conditions, but interpreting them is technically challenging.

Unlike written language, brain signals do not naturally divide into words or tokens.

When researchers attempt to process EEG data using conventional token-based approaches, they must decide how to divide continuous signals into smaller units. These divisions can be arbitrary and may not capture the underlying information effectively.

While developing a foundation model for brain signals with support from India’s Ministry of Electronics and Information Technology (MeitY) and the IndiaAI Mission, NeuroDX began exploring alternatives.

The team experimented with modelling the underlying meaning of information rather than depending entirely on predefined tokens.

The architectural ideas developed during this research eventually led the company to experiment with language models.

This research path ultimately produced Sphota, demonstrating how innovations developed for one scientific field can potentially find applications in another.

IndiaAI Mission Supports NeuroDX’s Frontier AI Ambitions

NeuroDX is among the startups selected to develop sovereign AI models under the IndiaAI Mission.

According to CEO Siddharth Panwar, the company received an allocation of 368 NVIDIA H200 GPUs for 18 months, alongside ₹12.5 crore in funding support, to develop a 20-billion-parameter multimodal EEG model.

This support highlights the importance of government-backed computing infrastructure for Indian AI startups.

Developing advanced AI models requires substantial computing resources, specialised engineering talent and access to large datasets. These requirements can create significant barriers for smaller companies competing with established international laboratories.

Panwar has argued that India should not abandon frontier AI development simply because domestic companies have fewer financial and computing resources than global competitors.

He believes Indian startups can pursue alternative technological approaches rather than directly replicating the architectures and strategies of established AI companies.

The company is also looking to raise more than $10 million to support its broader ambitions, although a completed funding round has not been announced.

Why Small AI Models Could Challenge Larger Systems

The development of Sphota highlights an important trend in artificial intelligence: improving intelligence per parameter rather than simply increasing model size.

Large AI models can require expensive GPU clusters, significant electricity consumption and substantial infrastructure investments.

Smaller models offer several potential advantages.

Lower Computing Requirements

A model with fewer parameters generally requires less memory to store and run, although actual inference costs also depend on architecture, precision, context length and serving infrastructure.

This could make smaller models attractive for startups and organisations that cannot afford large-scale GPU infrastructure.

Better Memory Efficiency

NeuroDX claims Sphota delivers approximately 4.5 times better memory efficiency than the comparable Liquid AI model.

This is particularly relevant for applications involving long conversations, where an AI system needs to maintain information from previous interactions.

However, the company has not yet published independent evaluations demonstrating how the improvement translates across different workloads.

Reduced Training Data Requirements

The company’s most notable claim is that Sphota achieves comparable performance using approximately 50 times less training data.

If independently validated, this could have important implications for AI development.

Training high-performing models requires large datasets, substantial computing capacity and considerable financial resources.

A more data-efficient architecture could help smaller research teams develop competitive models with fewer resources.

However, it remains unclear whether the reported efficiency gains will persist across different tasks and larger model sizes.

NeuroDX Continues to Develop AI for Healthcare

Although Sphota has brought attention to the company’s foundation model research, healthcare remains an important part of its work.

NeuroDX is developing EEG-based screening systems intended to help identify people who may be at risk of neurological conditions, including:

• Dementia.
• Epilepsy.
• Depression.

The company’s approach involves collecting brain signals using relatively inexpensive EEG equipment and processing them through AI models.

The objective is to help doctors at primary healthcare centres identify potential neurological risks and determine whether patients should be referred to specialists.

This could be particularly valuable in rural areas where access to neurologists and advanced medical facilities is limited.

Panwar said the healthcare models were undergoing evaluation and could be ready for deployment by the end of 2026 or early 2027, subject to government and healthcare partnerships.

The company has also collected labelled EEG data with ethical approvals and consent to support its research.

What Comes Next for Sphota?

NeuroDX is already working on a significantly larger version of its language model.

According to Panwar, the company is targeting a model containing between 1 billion and 2 billion parameters around December 2026.

The larger model will be an important test of whether Sphota’s reported efficiency advantages can be maintained as the architecture scales.

The company has also indicated that users could potentially access the larger model through a chatbot on its website.

For now, the Sphota architecture remains proprietary, meaning independent researchers cannot fully inspect its underlying design. The company says its brain foundation models are publicly available, while models developed with government support are intended to be open source.

This distinction is important because independent access and reproducible evaluations will be necessary to establish the architecture’s broader scientific and commercial significance.

The Bigger Picture

Sphota represents an interesting direction in India’s attempt to establish itself as a serious participant in frontier AI development. Instead of competing exclusively through larger models, bigger datasets and more powerful GPU clusters, NeuroDX is exploring whether alternative architectures can deliver better efficiency.

The reported results are encouraging, particularly the claims of 50 times less training data and 4.5 times better memory efficiency. However, the small benchmark difference between Sphota and Liquid AI’s comparable model, combined with the absence of independent verification, means the claims should be treated as preliminary rather than proof of a technological breakthrough.

If the approach continues to perform well at larger scales, it could provide Indian AI startups with another route to compete internationally despite limited computing resources. The larger opportunity is not simply building another language model but discovering architectures that make advanced AI more accessible and affordable.

Looking Ahead

The next major milestone for NeuroDX will be the development of its planned 1-billion to 2-billion-parameter model. Its performance will help determine whether the efficiency advantages observed in Sphota’s 240-million-parameter version can translate into larger and more capable systems. Independent benchmarking and broader developer access will also be important for validating the company’s claims.

For India’s AI ecosystem, Sphota highlights the importance of supporting unconventional research alongside conventional foundation model development. As the IndiaAI Mission expands access to computing infrastructure and funding, startups such as NeuroDX could contribute new approaches to AI architecture, healthcare technology and sovereign AI development. The ultimate test will be whether these innovations can move beyond promising early results and become widely adopted technologies.