Sphota AI model: NeuroDX claims 240M model edges Liquid AI

NeuroDX, an Indian start-up backed under the government's IndiaAI Mission, has unveiled Sphota, a small artificial intelligence (AI) language model with 240 million parameters, and says it beats a comparable model from the US company Liquid AI across about 11 benchmark tests. According to the company, Sphota scored an average of 52.4% on those benchmarks against 52.1% for Liquid AI's model, while using about 50 times less training data. The launch was reported in early November 2026. The claims have not yet been checked independently, and NeuroDX says it plans a larger model of 1 billion to 2 billion parameters around December.

Most attention in AI goes to huge models with hundreds of billions of parameters, the numbers inside a model that it adjusts while learning. Small models matter for a different reason: they can run on phones, laptops and cheaper servers, use less electricity and cost less to train. For India, where many users rely on mid-range phones and patchy internet, small models that work well could be more useful day to day than giant ones that only run in big data centres.

Sphota's main claim is not raw power but efficiency. If a 240-million-parameter model can reach the same scores as a comparable rival while being trained on far less data, that would lower the cost of building useful AI for Indian start-ups that do not have the budgets of global firms. Data efficiency is also important because good training data in Indian languages is limited.

The claims, however, need care. A benchmark is a standard set of test questions used to compare models, and small differences can come from how a test is run. The gap NeuroDX reports, 0.3 percentage points, is very small. One review of the claims noted that a gap this narrow, with no independent verification, means the results should be treated as preliminary. A research paper published in August 2026 that compared publicly benchmarked Indian foundation models also listed no standard benchmark results for NeuroDX at that time.

The launch adds to a busy week for Indian AI. On 3 November 2026, the IndiaAI Mission released three other home-grown foundation models, Soket AI Labs' EKA-1 and ZenteiQ's Dhīr 2B and Dhīr 5B. Together, these show the government's AI programme beginning to produce models of different sizes for different uses.

Peepals calculation: Sphota's reported average of 52.4% against 52.1% for Liquid AI's comparable model is a gap of 0.3 percentage points (52.4 − 52.1), or about 0.6% in relative terms (0.3 ÷ 52.1 ≈ 0.0058). The company's planned next model, at 1 billion to 2 billion parameters, would be about 4 to 8 times Sphota's size (1,000 ÷ 240 ≈ 4.2; 2,000 ÷ 240 ≈ 8.3). Our calculation from the reported figures.

Short-term: Developers and researchers can watch for NeuroDX's technical details and any independent tests that confirm or challenge the reported scores.

Long-term: If the efficiency claims hold up, Indian start-ups could build capable small models at lower cost, suited to phones and offline use. If they do not, the episode will add to calls for common, independent testing of Indian AI models.

Who is affected: Indian AI developers and start-ups; businesses looking for cheaper AI that runs on their own devices; the IndiaAI Mission, which backs the company; and users who could benefit from AI on low-cost phones.

What to watch next: NeuroDX's planned model of 1 billion to 2 billion parameters around December 2026, and any independent evaluation of Sphota.

Questions and Answers