The four key startups leading the race to build India's first LLM

With the government's support, work begins on foundational models for AI services

India will be ready with its first indigenously developed artificial intelligence large language model (LLM) in six to eight months, said Ashwini Vaishnaw, Union Minister for Electronics and Information Technology, on January 30 this year. 

The move was seen as India’s response to DeepSeek, an open-source LLM developed in China, reportedly at a fraction of the cost it took to create other models globally. (LLMs are AI programmes trained on vast amounts of text data to understand, generate, and process human language.) 

But Indian government officials say they are not merely reacting to events elsewhere. A plan to develop indigenous LLMs was already in the works, evident from the government accepting bids for more than 18,000 graphics processing units (GPUs) when the initial demand projection was just 10,000. 

“We have continued accepting bids from companies to procure and supply GPUs. If not for LLMs, where else could all these GPUs possibly be used?” a senior government official told Business Standard. 

As of May 31, the government had selected four startups to work on developing various LLMs: the first was Sarvam, followed by Soket AI, Gnani.ai, and Gan.ai. 

Sarvam, which was founded by Vivek Raghavan and Pratyush Kumar and is based in Bengaluru, has proposed to make three LLM versions: Sarvam-Large for advanced reasoning and generation, Sarvam-Small for real-time interactive applications, and Sarvam-Edge for compact on-device tasks. 

Soket AI, which has been in AI research for almost a decade, shifted its focus to high-performance computing and natural language processing in 2019, its founder and chief executive officer (CEO) Abhishek Upperwal told Business Standard. 

The Bengaluru-based company has proposed to develop India’s first open-source 120 billion parameter foundation model that will be optimised for the country’s linguistic diversity and will target the defence, health care, and education sectors. 

“We will be starting with 1-2 billion parameter models, test, and then scale up, or else we will waste the resources. In the process, we will release smaller models but they won't be as powerful,” said Upperwal. 

The trigger to build an indigenous LLM that understands Indian languages was OpenAI founder Sam Altman’s challenge that any model developed anywhere would always trail what ChatGPT had developed. While Upperwal initially wanted to conduct cutting-edge research work in India and build frontier models, the company pivoted to prioritising Indian languages. 

“That is also a motivation for the team to see how this intelligence emerges, what new can we do, and what ground-level problems can be solved with that intelligence? Ultimately, we are going to achieve the first AGI (artificial general intelligence) out of India,” said Upperwal. 

Gnani.ai, which is also based in Bengaluru, has proposed to build a voice-to-voice foundational model, capable of real-time speech processing and advanced reasoning. 

“The first LLM version, which will be ready within six months, will have 14 billion parameters, while the more advanced version, which could take up to a year, will have up to 70 billion parameters. We aim to produce near-instant autonomous voice conversations with very low latency,” said Ganesh Gopalan, Gnani’s co-founder and CEO. Gnani AI is also developing an LLM model capable of emotionally aware conversations.  

Gan.ai, which has proposed to create a 70-billion parameter multi-lingual foundation model, was founded by two Stanford University students Suvrat Bhooshan and Parth Sarthi. The startup aims to create an LLM with “super-human text-to-speech” capabilities that will surpass the current global models, according to government officials. The bootstrapped startup has not zeroed in on a location in India yet.