BharatGen's Ambitious AI Plans: From Trillion-Parameter Models to Domain-Specific Solutions

BharatGen's 'Recipe' for Building a Trillion Parameters Indic Model | AIM

The consortium insists sovereignty doesn't mean shutting the door on global players. BharatGen, the IIT Bombay-led consortium that bagged the biggest GPU allocation under the IndiaAI Mission, is now laying out the framework to achieve its most ambitious target yet -- a trillion-parameter model. The task is not just about scaling compute, but building the scaffolding India lacks -- data, talent, and what the group calls "recipes" for sovereign AI. "We picked this ambitious goal because we really want to move the needle on what's possible to build in India today," Rishi Bal, head of BharatGen, told AIM. "But this is not just about the models. It's about the entire ecosystem... it's a steep ramp." The numbers are staggering. BharatGen has secured 13,640 H100 GPUs and close to ₹1,000 crore in funding, the single-largest allocation in the co

BharatGen bets on small models to hit big AI goals - The Economic Times

An academic consortium based out of IIT Bombay, BharatGen last week secured the largest government support under the India AI Mission -- Rs 988.6 crore. Seven other entities were also selected under the mission to receive incentives for building foundational AI models. Government-backed AI initiative BharatGen will develop a suite of generic and domain specific small language models (SLMs) with sectoral applications, to build up a sovereign resource base for Indian firms as well as unlock enterprise value, its executive vice president Rishi Bal told ET. An academic consortium based out of IIT Bombay, BharatGen last week secured the largest government support under the India AI Mission -- Rs 988.6 crore. Seven other entities were also selected under the mission to receive incentives for building foundational AI models. The centre is banking on BharatGen to create foundational large language models (LLMs) and multimodal models, and make a mark globally as India's sovereign AI. But smaller models also remain a key goal. "We will work with Nasscom and push for a sovereign AI stack that companies looking for domain specific, effective small language models can access. We have already created domain specific SLMs, and will build on that," Ganesh Ramakrishnan, principal investigator at BharatGen, said. However, the road to developing SLMs rests on BharatGen's planned LLM with up to one trillion parameters, one of the key deliverables after the latest funding. A parameter is a variable that AI models learn from training data, and this ultimately determines the output based on the input data. "The ability to get up to that size (one trillion) will help in creating better, smaller, and more distilled models. We have already created SLMs for agriculture, ayurveda, legal, and finance sectors. A range of SLMs with different capabilities will be a key part of unlocking our enterprise potential," Bal explained. "There is a trajectory of small to large models, and vice versa. Some intermediate models will be better achieved when distilled from large models'', Ramakrishnan said. BharatGen is India's first indigenously developed multimodal LLM project for Indian languages, and is supported by the Department of Science & Technology under the National Mission on Interdisciplinary Cyber-Physical Systems. Multimodal models are designed to process, integrate, and analyse multiple 'modalities' of data simultaneously, such as text, images, audio, and video. The initiative is developing inclusive and efficient AI across 22 Indian languages. BharatGen's under-development LLMs will build on Param-1, a bilingual LLM with 2.9 billion parameters that it launched in May. This was pretrained on high-quality data from diverse Indian domains, across five trillion tokens in English and Hindi. A token is the basic building block of an LLM, being the smallest unit of data that an AI model processes, especially in natural language processing (NLP) and generative AI.