Sarvam AI: Building India's Sovereign AI Stack

Sarvam AI has emerged as one of India's most prominent AI companies. In April 2025, the Indian government selected Sarvam to build the country's sovereign large language model under the IndiaAI Mission — a government-backed initiative with roughly ₹10,372 crore (about $1.1 billion) in funding. The goal: an LLM built and deployed entirely in India, with strong Indian-language, reasoning, and voice capabilities.

A Growing Model Portfolio

Sarvam's product lineup now spans text, speech, translation, and vision. These are few models along with their use cases:

Sarvam-105B: The Flagship Model

Sarvam-105B is the company's most significant release to date, and it's especially notable for developers. Unveiled at the India-AI Impact Summit in February 2026, it's a 105-billion-parameter Mixture-of-Experts model — with roughly 9–10 billion active parameters per token — trained entirely from scratch on Indian soil, rather than fine-tuned from an existing foreign base model.

Key highlights:

1. 128K context window, with native support for all 22 official Indian languages
2. Open-sourced under Apache 2.0, with weights available on Hugging Face and AI Kosh, so startups and enterprises can deploy and modify it commercially without licensing fees
3. Built for reasoning, coding, document analysis, and agentic workflows
4. Powers Indus, Sarvam's own AI assistant for complex reasoning and multi-step tasks

Alongside Sarvam-105B, the company also released Sarvam-30B, a lighter, more efficient MoE model aimed at real-time conversational use cases — it powers Sarvam's conversational agent platform, Samvaad.

Why It Matters

Sarvam-105B represents one of the first fully domestically trained, large-scale open-source LLMs to come out of India, positioning the country as a genuine participant in frontier AI development rather than a downstream consumer of models built elsewhere. Combined with its speech, translation, and vision models, Sarvam is assembling a full-stack, India-first AI ecosystem — one built specifically for the way Indian languages are actually spoken, written, and mixed in everyday use.