Sarvam 30B
Sarvam AI · 2026-03-06 · 32.2B parameters
Sarvam 30B is Sarvam AI's 30B Mixture-of-Experts language model with approximately 2.4B active parameters, trained from first principles—including pretraining, SFT, and RL—in India as part of the IndiaAI Mission compute program and released under the Apache 2.0 license in March 2026. Its pretraining used 16 trillion tokens. It supports a toggleable thinking mode via `enable_thinking`, covers 22 Indian languages, and is designed for practical deployment balancing strong multilingual performance with efficient inference.