Government-backed BharatGen consortium is building one of the first non-profit population-scale large language models (LLM) in the country. This comes on the heels of Prime Minister Narendra Modi’s recent speech at the AI Summit in Paris, emphasising on India building its own LLM and foundational model considering the country’s diversity. The consortium includes IIT Bombay, IIT Kanpur, IIM Indore, and IIIT Madras. It is funded by the Department of Science and Technology (DST) of the Government of India. These India-centric foundational models are based on over 2 billion parameters for the text models for Hindi and English languages. There are 5 billion parameter models being used to build a diverse range of Indian context data for 13 Indian languages in total. The team is also using other methods to save on energy costs, like focusing on canonical representation of languages that involves focusing on the order of any word in a particular sentence. If successful, the LLM could be a worthy competitor to Chinese AI startup DeepSeek’s R1 AI model, which they claimed have built with an investment of $5.6 million and using only 2,000 GPUs, as compared to OpenAI’s GPT-4o that required $100 million to train the model.
