NeuroDX, an IndiaAI Mission-backed startup that began by building AI for brain signals, says its new 240-million-parameter language model, Sphota, has edged past a comparable model from US-based Liquid AI across roughly 11 benchmarks while using about 50 times less training data. The company, which is rebranding as Mannas, shared the preliminary results on 2 November 2026.

The figures are company-reported and have not yet been independently evaluated. Even so, they point to a different way of building language models, one that tries to capture meaning directly instead of relying mainly on predicting the next word.

What NeuroDX is claiming for Sphota

• Benchmark average: 52.4% for Sphota against 52.1% for the comparable Liquid AI model.
• Training data: comparable performance with around 50x less training data.
• Memory: about 4.5x better memory efficiency when holding long conversational context.
• Next step: a larger 1–2 billion parameter model targeted for around December 2026, possibly accessible through a chatbot on the company’s website.

CEO Dr Siddharth Panwar, who is also an assistant professor at IIT Mandi, has cautioned that the early efficiency gains may not carry over in the same proportion once the model is widely available.

From EEG signals to a new architecture

Sphota grew out of NeuroDX’s work on foundation models for electroencephalogram (EEG) signals, developed with support from MeitY and the IndiaAI Mission. Brain signals do not split neatly into tokens the way text does, and forcing them into tokens meant making arbitrary choices. The team began modelling the meaning carried in the signal instead, and found the approach worked for language as well.

The model’s name comes from sphota, the idea associated with the Indian grammarian-philosopher Bhartrihari that meaning is grasped as a whole rather than assembled word by word. In practice, the architecture is designed to recognise that two differently worded sentences describing the same event carry the same meaning, something conventional models learn only after training on very large amounts of data.

The company says its brain foundation models are publicly available and that models built with MeitY and IndiaAI support are intended to be open source, but the Sphota architecture itself will stay proprietary for now.

A sovereign AI bet on efficiency, not scale

NeuroDX is among the startups selected to build sovereign models under the IndiaAI Mission, and was allocated 368 Nvidia H200 GPUs for 18 months and ₹12.5 crore to develop a 20-billion-parameter multimodal EEG model. Panwar has argued that limited compute should not push India out of the frontier race, and that the better route is to pursue architectures that differ from the transformer path global labs have already invested in. The company is also looking to raise more than $10 million, though no completed round has been announced.

The push sits alongside other compute-light Indian efforts, such as Murai Labs’ Agaram 2.5B Tamil-first model, and the government’s work to widen access through the IndiaAI Mission’s 10,000-GPU tender.

Healthcare remains the anchor

The company is also building low-cost EEG screening that could help MBBS doctors at primary health centres and in villages flag people at risk of dementia, epilepsy and depression and refer them to specialists. Those models are under evaluation and could be ready for deployment by the end of 2026 or early 2027, subject to work with government and healthcare partners.

Why it matters for CIOs

If small, data-efficient models can match larger ones, enterprises could run capable AI on less hardware, at lower cost and closer to sensitive data. The real test is whether Sphota’s numbers hold up under independent benchmarking and at the 1–2 billion parameter scale. Technology leaders should treat the claims as an early signal worth tracking, alongside India’s broader sovereign AI push.

Key takeaways

• NeuroDX (now Mannas) says its 240M-parameter Sphota model averages 52.4% on about 11 benchmarks, ahead of a comparable Liquid AI model at 52.1%.
• It claims about 50x less training data and 4.5x better memory efficiency; the results are preliminary and not independently verified.
• A 1–2 billion parameter version is targeted for around December 2026.

Source: NeuroDX (Mannas) CEO Dr Siddharth Panwar, as reported on 2 November 2026; company website.

Be a part of Elets Collaborative Initiatives. Join Us for Upcoming Events and explore business opportunities. Like us on Facebook , connect with us on LinkedIn and follow us on Twitter.

"Exciting news! Elets technomedia is now on WhatsApp Channels Subscribe today by clicking the link and stay updated with the latest insights!" Click here!