Sarvam 105B
Sarvam AI · 2026-03-06 · 106B parameters
Sarvam 105B is Sarvam AI's 105B Mixture-of-Experts language model with approximately 10.3B active parameters, trained from scratch in India using compute provisioned through the IndiaAI Mission and released under the Apache 2.0 license in March 2026. Its pretraining used 12 trillion tokens before supervised fine-tuning and reinforcement learning. It supports 22 Indian languages and is optimized for complex reasoning, mathematics, coding, and agentic tasks across a 128K-token context window.