A list of Indian-built AI models published by NVIDIA on 17 February 2026 gives a rough picture of how the country’s home-grown models were sized at that time, according to the company’s blog on the IndiaAI Mission. The blog calls its list a representative sample and not a complete catalogue, and IndiaX24 has not checked later releases.
BharatGen is listed with a 17-billion-parameter mixture-of-experts model. A mixture-of-experts design activates only part of a model’s parameters for each input, which is general background and not a detail given in the blog. The blog does not say which languages the model supports.
Sarvam.ai is listed with models of 3 billion, 30 billion and 100 billion parameters across 22 Indic languages, according to NVIDIA. IndiaX24 has reported separately on Sarvam AI’s valuation of 1.5 billion dollars and on a survey in which 63 per cent of investors said Indian AI might be overvalued. The 100-billion-parameter model is about 33 times the size of the 3-billion one, a ratio calculated by IndiaX24.
Gnani.ai is listed with a 14-billion-parameter speech-to-speech model, and NVIDIA says the company achieved a 15 times reduction in inference costs, according to the blog. Inference is the running of a trained model to answer requests, and a 15-fold cost cut would be significant for voice services, but the blog does not explain how it was measured or against what baseline.
Chariot is listed with an 8-billion-parameter text-to-speech model, and Tech Mahindra with an 8-billion-parameter foundation model for Indian languages, according to the blog. Together the five developers’ main models in the list add up to 17, 100, 14, 8 and 8 billion parameters at the top of their ranges, a sum of 147 billion that IndiaX24 offers only as an arithmetic illustration, since the models are separate and are not combined.
Parameters are a rough measure of size and not of quality. A larger model is not necessarily better for a given task, and a small model trained on good data can outperform a larger one in a narrow domain, which is general background. The blog gives no benchmark scores, so IndiaX24 does not rank the models.
The models run on NVIDIA hardware, which is why the company lists them. The blog’s account of the compute pillar names Yotta, L&T and E2E Networks, as IndiaX24 reports in a separate article. The same blog says that more than 4,000 Indian AI start-ups are in NVIDIA’s Inception programme and names venture firms including Peak XV, Z47, Elevation Capital, Nexus Venture Partners and Accel India as partners.
A new training dataset is also relevant. NVIDIA and AI4Bharat released a corpus of 2.8 trillion tokens in 22 languages on 16 October, as IndiaX24 reports, which could be used by developers such as these, though the reports reviewed do not say whether any of the listed companies will use it.
The limits of the list should be clear. It comes from a company with a commercial interest, it is eight months old, and it describes models that were announced and not necessarily released for public use. IndiaX24 has not tested any of them.
IndiaX24 will report on model releases and independent benchmarks as they appear.
Image: A technician at a server rack (file photo; a stand-in image). Photo by Derrick Coetzee from Berkeley, CA, USA, licensed under CC0, via Wikimedia Commons.
Leave a Reply