Even as Sarvam AI gets the government mandate to build a sovereign large language model from the ground up, AI experts want it to be culturally relevant, contextual to local history and traditions, and proficient in native languages.

“The benchmark is not good enough for local needs. It must strive to be among the best globally,” Kashyap Kompella, an AI industry analyst, told businessline.

“Sovereign LLMs must be good enough to be preferred by the private sector, not just government or subsidised users. Countries can leverage both Western LLMs and local LLMs based on the respective use cases,” he said.

On Saturday, the India AI Mission, which called for applications from Indian entities to build native foundational models, shortlisted Sarvam AI, a Bengaluru-based AI company, for developing the maiden sovereign foundational model.

“Building India’s sovereign model from the ground up is a crucial step toward Atmanirbhar Bharat. The model will be fluent in Indian languages, designed for voice, capable of reasoning, and ready for secure, population-scale deployment,” Sarvam AI said after bagging the mandate.

It will get a compute of 4,000 Graphics Processing Units (GPUs) for six months to build the 70-billion parameter model.

Call for policies

V Srinivasa Rao, former Chief Digital Officer of Tech Mahindra and Founder of CogniSpark, which offers consulting in agentic AI, called for the preparation of proper policies that govern the native LLMs. 

“They (the government) should come out with regulations. The datasets are required to be audited for various cultural, religious, regional, and linguistic sensitivities. If the datasets are not audited, it could result in problems when the native LLMs churn outputs with biases,” he pointed out.

“All States and regions should be consulted to understand the nuances. A Task Force should be set up to take care of ethical, responsible, and explainable AI,” he said.

Making a strong pitch for native LLMs, Kashyap felt that Western-trained LLMs often miss context, sensitivities, traditions, humour, and social norms specific to non-Western societies. 

“Local population is not proficient in English and poor LLM performance in local languages limits utility and creates barriers to entry. A local LLM can better understand and reflect societal values, folklore, and region-specific challenges,” he said.

“LLMs trained predominantly on Western datasets reinforce a Western-centric worldview. Local LLMs help reclaim historical accounts from a local perspective, rather than filtered through Western academic lenses,” he pointed out.

Published on April 27, 2025