Originals · Startup · Startups & Fintech
NeuroDX says its 240M Sphota model edges a Liquid AI rival with far less training data
The IndiaAI Mission backed startup, which began with AI for brain signals, reports a 52.4% benchmark average against 52.1% for a comparable Liquid AI model. The numbers are its own and not yet independently checked.
NeuroDX, an Indian startup backed under the IndiaAI Mission, shared preliminary results on 2 November 2026 for Sphota, a small language model with 240 million parameters. According to the company, Sphota averaged 52.4% across about 11 benchmarks, against 52.1% for a comparable model from US based Liquid AI. Analytics India Magazine and Elets CIO both reported the figures, and both stress that they are company reported and have not been independently evaluated.
The bigger claims are about efficiency. CEO Siddharth Panwar, who is also an assistant professor at IIT Mandi, says Sphota reaches similar performance with about 50 times less training data and holds long conversational context with about 4.5 times better memory efficiency. Panwar himself cautioned that these early gains may not carry over in the same proportion once the model is widely available.
The model grew out of NeuroDX's original work on foundation models for EEG signals, the electrical readings taken from the scalp. Brain signals do not split neatly into tokens the way text does, so the team began modelling the meaning carried in a signal instead of chopping it into pieces, and found the approach also worked for language. The name comes from the idea, linked to the grammarian Bhartrihari, that meaning is grasped as a whole rather than word by word. Elets reports that the Sphota architecture will stay proprietary for now.
NeuroDX is among the startups selected to build sovereign models under the IndiaAI Mission. Elets reports it was allocated 368 Nvidia H200 GPUs for 18 months and Rs 12.5 crore to build a 20 billion parameter multimodal EEG model, and that it is looking to raise more than $10 million, though no completed round has been announced. The company is targeting a larger model of 1 to 2 billion parameters around December 2026.
Why it matters: India's model builders have limited compute next to global labs, and a credible small, data efficient architecture would be a different way to compete. Peepals notes that the reported margin is only 0.3 percentage points and that a research survey in August 2026 found no public standard benchmark results for NeuroDX, so independent evaluations will decide how much weight the claim deserves.
