NeuroDX's 240M-parameter Sphota scores 52.4% vs Liquid AI's 52.1% across about 11 benchmarks, company claims

NeuroDX, an Indian startup backed by the IndiaAI Mission, says its new 240-million-parameter language model, Sphota, scored 52.4% on average across about 11 benchmarks, compared with 52.1% for a comparable model from Liquid AI. The figures come from the company's own reported results, as covered by Analytics India Magazine, and have not yet been established through independent public evaluations.

Key claims

• Parameters: 240 million
• Average benchmark score: 52.4% for Sphota versus 52.1% for Liquid AI's model, across about 11 benchmarks (company-reported)
• Training data: NeuroDX CEO Siddharth Panwar claims comparable performance with 50 times less training data. He says the figure compares the training data used by Sphota and Liquid AI's model to reach similar benchmark performance.
• Memory efficiency: Panwar claims 4.5 times better memory efficiency, which relates to how much conversational context a model can retain under comparable conditions.
• Context window, pricing, training cutoff: not yet disclosed

Panwar cautioned that the efficiency gains observed so far may not translate directly into the same numbers when the model becomes widely available. The company's figures are reported comparisons, not independently verified results across all models and workloads.

Architecture built around meaning

Sphota grew out of NeuroDX's work on brain research. The company initially built AI models that interpret electroencephalogram (EEG) signals, developing a brain foundation model with support from the Ministry of Electronics and Information Technology (MeitY) and the IndiaAI Mission. EEG signals lack the natural divisions of written language, so splitting them into tokens requires arbitrary decisions. The team explored modeling meaning directly instead, and the resulting architecture proved applicable to language.

Most language models divide text into tokens and learn by predicting what comes next. NeuroDX says Sphota instead makes identifying underlying similarities between differently worded descriptions of the same event a design principle. The name comes from a linguistic concept associated with Bhartrihari, who argued that meaning is apprehended as a whole.

The company says the approach could help smaller models develop certain capabilities with less training data. Whether the gains hold across a wider range of tasks and larger models remains to be established. NeuroDX plans to keep the Sphota architecture proprietary for now.

Next model and rebrand

NeuroDX has worked on Sphota for only a few months with a small team. It is developing a larger model of between 1 billion and 2 billion parameters, which Panwar targets for release around December. He said users could potentially access it through a chatbot on the company's website.

The company is also rebranding as Manas AI, reflecting a shift from a healthcare focus to foundation models. It remains involved in neurological healthcare. Panwar said the company's brain foundation models are publicly available, while models developed with MeitY and IndiaAI Mission support are intended to be open source.

Compute and funding

NeuroDX is among the companies selected to develop sovereign AI models under the IndiaAI Mission. Panwar said it was allocated 368 H200 GPUs for 18 months and ₹12.5 crore to develop a 20-billion-parameter multimodal EEG model. The company is looking to raise more than $10 million. No funding round has been completed.

What this means

The reported gap between Sphota and Liquid AI's model is 52.4% versus 52.1% on an average across about 11 benchmarks. The more notable claims concern training data (50x less) and memory efficiency (4.5x), and those rest on company comparisons that outside evaluators have not yet tested. Panwar himself says the efficiency numbers may shift once the model is widely available.

The 1B to 2B parameter model targeted for around December is the next test, since the company says it remains to be established whether Sphota's gains persist at larger scales. Until then, Sphota is an architecture claim at 240 million parameters that outside researchers cannot inspect, because the company is keeping it proprietary.