The aspiration at Sarvam, as the word Sanskrit meaning ‘all’ suggests, is to achieve in AI (Artificial Intelligence) something that encompasses everybody. That’s precisely what Vivek Raghavan and Pratyush Kumar set out to do when they founded the Bengaluru-based startup in July2023. “We want to build GenAI that a billion Indian can use,” they would often say in their multiple public interactions.
The moment to finally walk the talk has finally arrived for the two scholars with a wealth of experience. With the government choosing Sarvam to build India’s first sovereign foundational AI model, the two founders now step into a new, chapter of their entrepreneurial journey that will bring them closest to the vision that drove Sarvam’s creation: building population-scale AI solutions.
What ignited their vision was the emergence of DeepSeek, which disrupted the AI landscape that was largely shaped by western tech behemoths. “DeepSeek proved that one doesn’t need billions to build a competent model,” the duo had said in a media interview.
Before founding Sarvam, Raghavan and Kumar had worked together at AI4Bharat, a research initiative by IIT Madras to develop open-source Indian language AI. The initiative developed a wide range of datasets, tools, and models focusing on natural language understanding, automatic speech recognition, transliteration-converting text between scripts of Indian languages and English, and speech synthesis.
AI4Bharat started in 2020 as a collaboration between IIT Madras and Nandan Nilekani’s ekStep foundation, where Raghavan was working as the Chief AI Evangelist. In his years at the foundation, which also acted as a knowledge partner for the National Language Translation Mission – Bhashini — Raghavan contributed to the development of Bhashaverse, an application for speech-to-speech translation across 11 Indian languages and text-to-text translation in all 22 Indian languages. This work eventually culminated in the release of the Bhashini app. Raghavan has parallely spent more than a decade at UIDAI, building India’s digital infrastructure, such as Aadhaar.
While working for AI4Bharat, Kumar joined IIT Madras as an adjunct faculty member in late 2021. A PhD graduate from ETH Zürich and an IIT Bombay alumnus, Kumar has previously worked as a research scientist at IBM and later at Microsoft. A few years later, they co-founded Sarvam AI with a vision to build a generative AI stack from scratch. Peak XV Partners and Lightspeed Venture Partners were among the first investors who cut them a $12.6 million seed cheque and then later made follow-on investments in their $41 million Series A round in December 2023.
Because building a large language model from scratch is enormously resource and capital-intensive, Sarvam started by taking open-source models such as Meta’s Llama 2 and fine-tuned them on Indian datasets to make them better at understanding and generating text in Indian languages and contexts.
In August last year, Sarvam released its full-stack Gen AI platform with a suite of products, including custom AI voice agents for businesses (deployable via telephone, WhatsApp, or in-app and available in 10 Indian languages), Sarvam 2B (India’s first foundational, open-source, small Indic LLM, trained from scratch on an internal dataset of four trillion tokens), and Shuka 1.0 (an audio extension on the Llama 8B model to support Indian language voice in and text out).
Later that year, in October, the startup launched Sarvam-1, a 2-billion-parameter model, trained on four trillion tokens curated by Sarvam on NVIDIA H100 Tensor Core GPUs. Its custom tokeniser is up to four times more efficient than leading English-trained models on Indian language text, according to its blog post from last year.
Besides English, Sarvam-1 supports 10 Indian languages: Bengali, Gujarati, Hindi, Marathi, Malayalam, Kannada, Oriya, Tamil, Telugu, and Punjabi.
Now, the task at hand is to build a 70-billion parameter AI model optimised for voice, reasoning, and fluency in 22 Indian languages, under the government’s Rs 10,000-odd crore IndiaAI Mission. The size of an AI model is often described by the number of parameters it has. More parameters generally mean that the model can learn more complex patterns, but it also requires more data and computational power to train effectively. For this, the government will provide Sarvam with access to 4,096 Nvidia H100 GPUs for six months from the IndiaAI Mission’s AI compute portal.
More recently, Sarvam has been working with NITI Aayog to develop a pilot for an Enterprise Reasoning Engine (ERE) on the National Data and Analytics Portal (NDAP), aiming to enhance the platform’s data accessibility and usage. Sarvam has also deployed AI solutions to enhance the user experience of Aadhaar services, including voice-based interactions.
The next six months are obviously crucial for the duo as they work to develop a foundational AI model that balances both public impact and commercial viability. But this venture also marks a defining moment for India as there was a crying need for an AI-first company that builds from the ground up — not just tinkering at the edges of innovation but owning the full stack.