Sarvam-105B is a monumental achievement in the 2026 AI landscape, serving as India’s first large-scale sovereign foundation model. Launched in February 2026 under the government’s IndiaAI Mission, this 105-billion parameter model was trained entirely from scratch on domestic compute infrastructure. Unlike earlier efforts that fine-tuned foreign models, Sarvam-105B represents a “ground-up” approach, utilizing a Mixture-of-Experts (MoE) architecture that activates only ~9 billion parameters per token, making it both highly efficient and environmentally sustainable.

The model is specifically engineered for complex reasoning, coding, and enterprise-grade workflows. It features a massive 128,000-token context window, allowing it to digest entire legal filings or financial reports in a single prompt. Benchmarks show it outperforming models six times its size, including certain evaluations where it surpassed China’s DeepSeek-R1. While global giants like Llama 3.1 are larger in raw parameter count, Sarvam-105B’s competitive edge lies in its native understanding of all 22 official Indian languages and code-mixed formats like “Hinglish,” making it the definitive choice for Bharat-scale deployment.