An Indian startup says its 240-million-parameter Sphota language model slightly outscored a comparable Liquid AI model across about 11 benchmarks, while using far less training data. The result could point to a different route to efficient AI, but it remains the company’s reported comparison, not an independently verified win.
Watch Desk analysis
What happened
NeuroDX chief executive Siddharth Panwar told Analytics India Magazine that Sphota scored 52.4% against 52.1% for the comparable Liquid AI model. He also claims Sphota used 50 times less training data to reach similar performance and is 4.5 times more memory-efficient. The article says the results are preliminary and have not been established through independent public evaluations.
The model grew out of NeuroDX’s work on AI for brain signals, where conventional text-token approaches did not map neatly onto EEG data. The company says Sphota is designed to identify underlying meaning rather than rely primarily on predicting the next word. NeuroDX is developing a larger model, targeting a release around December, and plans to rebrand as Manas AI.
Why it matters
If the reported gains hold up beyond these early comparisons, a smaller model that needs less data and memory could make some AI development and deployment more accessible. The architecture’s origins in EEG research also make this more than another routine leaderboard claim: the company says an approach developed for brain signals led it towards a different way of building language models.
But the evidence described so far is narrow. Results across about 11 benchmarks do not establish broad capability, and the reported efficiency figures depend on comparisons that have not been independently tested in public. The larger model will be a more revealing test than a close early score.
Our read
Sphota is an intriguing research direction, not yet a challenger crowned by the numbers. The worthwhile question is whether the model’s reported efficiency survives wider, like-for-like evaluation as it scales. Until then, 50 times less training data is a striking company claim, not a general rule for building language models.
What to watch
- Whether NeuroDX publishes benchmark details and enough information for independent replication.
- Whether the planned larger model preserves the reported performance and efficiency gains.
- Whether the rebrand and product plans bring Sphota to users, and on what terms.
Discussion spark: What evidence would convince you that Sphota’s reported efficiency gains come from its architecture rather than from the particular benchmarks and comparison chosen?
Sources and evidence
- India’s Frontier AI Race Gets a 240M Model That Beats Liqui – Analytics India Magazine (2 November 2026, 07:28 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.