Sarvam AI Charts Full-Stack Future With Model, Apps

On 26 April, IIT Madras-incubated artificial intelligence startup Sarvam became the first to receive subsidized access to processing chips from the Union government to build Indian AI models.

- Shouvik Das

On 26 April, IIT Madras-incubated artificial intelligence startup Sarvam became the first to receive subsidized access to processing chips from the Union government to build Indian AI models. Now, after facing conjecture and criticism over government support and the startup's innovation route, its founder and investors believe that the way forward for India's most well-funded AI startup is to take a leaf out of the Big Tech playbook.

"We're not just a model builder," Pratyush Kumar, co-founder of Sarvam, told Mint on the sidelines of a gathering in Bengaluru last week. "We have an application layer too, and we want to be focused on both (the model and the layer)."

The startup has faced criticism after it released its AI model Sarvam-M last month. While some hailed the 24-billion-parameter model for its native support for 10 Indian languages, others criticized it for not being a foundational model, but being based on a foundational model made by a French AI startup, Mistral.

Kumar and his investors, however, do not believe that the criticism has been fair. "It's hard to quantify how much of our efforts we are devoting to foundational models versus building applications—the models will be key because this will give us the foundational capability. The government funding is just incidental, but it will not define the way forward for Sarvam. Eventually, we're building a full-stack, deep-tech firm out of India," he said.