Krux
February 22, 2026
India's Sarvam AI Unveils Models for 22 Languages
Published: February 22, 2026 at 3:03 PM
Updated: February 22, 2026 at 3:03 PM
100-word summary
Sarvam AI launched two large language models, Sarvam-30B and Sarvam-105B, trained for all 22 Indian languages using a mixture-of-experts design with up to 128,000-token context windows. The company also debuted Vikram, an offline multilingual assistant that runs on feature phones, cars, and wearables for real-time voice conversations in local languages. Demos showed Hindi and Punjabi interactions for government-scheme guidance. Built with Nvidia and Yotta infrastructure under India's sovereign IndiaAI Mission, the models aim to challenge frontier AI with cost-efficient, edge-optimized reasoning and mark a major step toward homegrown, multilingual AI infrastructure.
What happened
Sarvam AI launched two large language models, Sarvam-30B and Sarvam-105B, trained for all 22 Indian languages using a mixture-of-experts design with up to 128,000-token context windows. The company also debuted Vikram, an offline multilingual assistant that runs on feature phones, cars, and wearables for real-time voice conversations in local languages. Demos showed Hindi and Punjabi interactions for government-scheme guidance.
Why it matters
Built with Nvidia and Yotta infrastructure under India's sovereign IndiaAI Mission, the models aim to challenge frontier AI with cost-efficient, edge-optimized reasoning and mark a major step toward homegrown, multilingual AI infrastructure.