India's Sarvam and Krutrim build frugal AI models for local needs

Frugal AI for India's specific constraints

In response to India's linguistic diversity and infrastructure limitations, several Indian AI initiatives are building lightweight, cost-effective models that differ from compute-heavy Silicon Valley approaches. AI4Bharat, launched at IIT Madras in 2020, focuses on AI tools tailored for Indian languages and real-world constraints, building systems that run on low-end smartphones and low bandwidth networks.

Sarvam AI's approach and models

Sarvam AI, co-founded by Vivek Raghavan and Pratyush Kumar, develops full-stack AI solutions for India's diverse linguistic and cultural landscape. The company builds on frugal design principles similar to those used in Aadhaar and the Unified Payments Interface.

Key models include:

• SarvamM: A 24-billion parameter large language model trained across 10 Indian languages
• Sarvam 2B and Sarvam-M: Models fine-tuned for medical reasoning and symptom triage in local languages

Practical applications in healthcare and education

In healthcare, Sarvam AI deploys voice-enabled, multilingual conversational agents that allow rural patients to access medical advice through WhatsApp and low-bandwidth interfaces. These systems can summarize patient notes, offer diagnostic guidance, and prioritize cases without requiring high-end devices or constant internet.

In education, Sarvam's models enable vernacular learning assistants capable of understanding codemixed queries and delivering personalized instruction in students' mother tongues. These lightweight, optimized models adapt lessons in mathematics and programming to regional educational contexts.

Technical and strategic approach

Sarvam AI operates as a for-profit startup, believing significant investment and market competition are needed to scale AI's impact across India. The company's strategy involves open-sourcing AI models and collaborating with Indian enterprises to build domain-specific solutions. Their goal is to bring generative AI to 800 million Indians with smartphones.

The approach emphasizes sovereign AI models that respect data privacy and cultural nuances, addressing India's 22 official languages and over 1,600 dialects where global English-centric models fall short.