Sarvam AI launches India's sovereign LLM 'Sarvam-M', claims edge over LLaMA-3, Gemma 3 on key benchmarks

In a big development in India's AI space, Sarvam AI, the AI startup selected under the government's IndiaAI Mission to build India’s sovereign Large Language Model (LLM), has launched its flagship LLM called Sarvam-M. The LLM is a 24-billion multilingual, hybrid-reasoning, text-only language model built on top of Mistral Small, a compact yet efficient open-weight AI model developed by Mistral AI, a French AI company known for its high-performance language models.

The flagship AI model developed by Bengaluru-based Sarvam AI achieves a new benchmark across a range of Indian languages, math, and programming tasks, for a model of its size. Sarvam AI says Sarvam-M is built for versatility and designed to support a wide range of applications, including conversational agents, translation, educational tools, and so much more.

"This is the first in a series of contributions as we help build out the Sovereign AI Ecosystem in India." The company has said it will come with more updates in the coming weeks. "Towards Building a Sovereign AI Ecosystem in India, we plan to have regular model drops and share our detailed technical findings."

Sarvam-M is open-sourced and accessible via Hugging Face, with APIs provided for developers. "Sarvam-M represents an important stepping stone on our journey to build Sovereign AI for India. More to come," said Sarvam Co-founder Vivek Raghavan said on X.

In a technical blog released during the launch of the AI model, the Sarvam AI team shares explorations on post-training and inference optimisation of an open pre-trained model to create a cutting-edge hybrid reasoning model specialised for Indic languages.

In terms of performance benchmarks, the AI model shows 20% average improvement over the base model on Indian language benchmarks, 21.6% enhancement on math-related tasks, and 17.6% improvement in coding benchmarks.

Performance gains are even more impressive at the intersection of Indian languages and mathematics, with an outstanding +86% improvement in romanised Indian language GSM-8K benchmarks, highlighting its capability in handling code-mixed language scenarios.

The current version comes a month after Sarvam AI unveiled a new speech AI model Bulbul, which is offers natural, familiar speech in 11 Indian languages, with authentic accents that sound just like India.

Sarvam says the advanced Sarvam-M outperforms Llama-4 Scout on most benchmarks, and is comparable to larger dense models like Llama-3.3 70B, and models like Gemma 3 27B, which are pre-trained on significantly more tokens. One area where Sarvam has room for improvement, says the company, is "knowledge related benchmarks in English", such as MMLU where Sarvam-M drops about 1% points over the baseline model.

Key Features of Sarvam-M:

Hybrid Thinking Mode: Sarvam-M is a single versatile model supporting both "think" and "non-think" modes. Use the think mode for complex logical reasoning, mathematical problems, and coding tasks, or switch to non-think mode for efficient, general-purpose conversation.

Advanced Indic Skills: The model is specifically post-trained on Indian languages alongside English, a characteristics that authentically reflects and emphasises Indian cultural values.

Superior Reasoning: Outperforms most similarly-sized models on coding and math benchmarks, demonstrating exceptional reasoning abilities.

Seamless Chatting: The model offers full support for both Indic scripts and romanised versions of Indian languages, providing a smooth and accessible multilingual conversation experience.

Founded in July 2023 by Vivek Raghavan and Pratyush Kumar, Sarvam aims to make Generative AI accessible at scale in India. In April 2025, under the IndiaAI Mission, the Centre selected Sarvam to build India’s sovereign LLM. Sarvam will receive dedicated compute resources to build an indigenous foundational model from scratch.