Krux
February 21, 2026
India's Sarvam AI Drops Two Homegrown LLMs, Goes Open-Source
Published: February 21, 2026 at 3:43 AM
Updated: February 21, 2026 at 3:43 AM
100-word summary
Sarvam AI just unveiled two made-in-India language models at the India AI Impact Summit: Sarvam-30B for real-time chat and Sarvam-105B for complex reasoning. Both use mixture-of-experts architecture, activating only some parameters to stay efficient while handling massive context windows (32K and 128K tokens). Trained from scratch on trillions of tokens across Indian languages using government-backed compute, they rival global models but understand local languages better. Sarvam plans to open-source both and launched Indus, a consumer chat app powered by 105B. This could make advanced AI accessible across India's linguistic diversity, challenging Western-dominated models with sovereign, multilingual alternatives.
What happened
Sarvam AI just unveiled two made-in-India language models at the India AI Impact Summit: Sarvam-30B for real-time chat and Sarvam-105B for complex reasoning. Both use mixture-of-experts architecture, activating only some parameters to stay efficient while handling massive context windows (32K and 128K tokens). Trained from scratch on trillions of tokens across Indian languages using government-backed compute, they rival global models but understand local languages better. Sarvam plans to open-source both and launched Indus, a consumer chat app powered by 105B.
Why it matters
This could make advanced AI accessible across India's linguistic diversity, challenging Western-dominated models with sovereign, multilingual alternatives.