Sarvam, a Bangalore-based startup, has been assigned the task for building an LLM for India. Sarvam in Sanskrit means ‘all’, and it implies that the proposed LLM will encompass everybody.
The startup was founded by Vivek Raghavan and Pratyush Kumar in July 2023. They had in mind to create an LLM that can be used by a billion people. The arrival of DeepSeek, the economical Chinese model that was competent enough, fueled the ambitions of Sarvam founders.
Previously, Sarvam founders worked for AI4Bharat, a research initiative by IIT Madras to develop open-source Indian language AI. Apart from NLP, the model should have speech recognition capability, transliteration capability, converting texts of Indian scripts to English and speech synthesis ability.
AI4Bharat was a collaboration between IIT Madras and EkStep of Nandan Nilkeni. Raghavan worked here as chief AI evangelist.
At the same time, Raghavan worked as a knowledge partner for National Language Translation Mission — Bhashini. Raghavan facilitated the development of Bhasaverse, an app to do speech-to-speech translation across 11 Indian languages and text-to text translation in all 22 Indian languages. The work ultimately led to Bhashini app. Raghawan had also spent more than a decade simultaneously at Aadhar.
Kumar too worked for AI4Bharat and joined as an adjunct faculty at IIT Madras in 2021. Kumar had a doctorate from ETH Zurich and was an IIT Bombay alumnus. He had a research stint at IBM and Microsoft. He was a co-founder of Sarvam AI that proposed to build an indigenous LLM from scratch. The initial funding came from Peak XV and Lightspeed Venture.
The building of LLM is resource-intensive process. They initially worked with open-source models and fine-tuned them on Indian datasets. Later, they developed voice agents. They launched Sarvam-1, a 2-billion parameter model, trained on four trillion tokens in October 2024. It supports apart from English 10 Indian languages.
The government has now assigned them a task to build a 70-billion parameter AI model optimized for voice reasoning and fluency in 22 Indian languages. More parameters enable the model to learn more complex patterns. It also requires more data and computational power. The government will provide Sarvam access to 4096 Nvidia H100 GPUs for six months.
Sarvam is also working with NITI Aayog to develop a pilot for Enterprise Reasoning Engine (ERE) on the National Data and Analytical Portal (NDAP) to enhance portal’s data accessibility and usage.
Sarvam has also deployed AI solutions to enhance user experience of Aadhar services, including voice-based interactions.
The next six months are critical for this pair of Raghavan and Kumar as they try to develop a foundation model. It makes a defining moment for India since an Indian startup is not tinkening at the edges of AI but owning the full stack.