India's approach to AI looks nothing like Silicon Valley's. Startups Sarvam AI and Krutrim are building what they call "sovereign AI," small and cheap models built for local languages and infrastructure. These run on basic smartphones with spotty internet. You can't serve 800 million Indians with English-only models that need expensive hardware.
Sarvam AI, co-founded by Vivek Raghavan and Pratyush Kumar, grew out of the AI4Bharat initiative at IIT Madras. Their models include Sarvam 2B and SarvamM, a 24-billion parameter model trained across 10 Indian languages. These are deployed for real healthcare work right now. Voice-enabled agents let rural patients get medical advice, book appointments, and consult doctors through WhatsApp. The models handle medical reasoning and symptom triage in local languages, no high-end phone required.
The team hit a wall. A Hindi sentence required three to four times more tokens than the same sentence in English. Every interaction cost more. "The same question, when asked in English, costs one-fifth of what it costs in an Indian language," Raghavan explained. Sarvam built better tokens and higher-quality datasets to close that gap.
Krutrim, launched by Ola Cabs co-founder Bhavish Aggarwal in 2023, pursues similar goals with its own models. And the OpenHathi project takes a different approach, adapting existing open-source models like Meta's LlaMA and Mistral for Indian languages rather than training from scratch. The results land on Hugging Face, ready for any developer to pick up and use.