Sarvam, the startup selected for building India’s foundational LLM under the IndiaAI Mission, has unveiled Sarvam-M, a 24-billion parameter open-weights hybrid language model built on top of Mistral Small.
The model sets new benchmarks in Indian language understanding, mathematics, and programming tasks for its size.
Designed for broad application, Sarvam-M is optimised to power use cases such as conversational AI, machine translation, and educational tools. It is now accessible through the Sarvam API.
In a technical blog accompanying the launch, the team detailed their approach to fine-tuning, reinforcement learning, and deployment optimisation. The model is heavily focused on coding and maths.
“We found the base Mistral Small model could be significantly improved in Indian languages,” reads the paper behind the reason for choosing Mistral Small.
Sarvam chose to generate approximately one-third of the training samples in Indic languages. Specifically, 30% of the coding, math, and reasoning prompts—and 50% of the remaining prompts—were translated into Indian languages.
Among these, Hindi accounted for 28% of the Indic data, while nine other languages—Bengali, Gujarati, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, and Telugu—each made up 8%. “Together these 10 languages represent the ‘mother tongue’ or first language of over 70% of the Indian population, i.e., about a billion people.”
This release marks the first of several initiatives Sarvam plans to roll out in support of building a Sovereign AI ecosystem in India. More updates are expected in the coming weeks as was announced by the company earlier.
Speaking on the launch, co-founder Vivek Raghavan, posted on X, saying, “Sarvam-M represents an important stepping stone on our journey to build Sovereign AI for India.”
Earlier this month, Sarvam AI unveiled Bulbul, a new speech AI model that supports 11 Indian languages with natural, region-specific accents. It offers authentic, human-like voice interactions.
Under the IndiaAI Mission, Sarvam AI had proposed the development of a 70-billion parameter multimodal AI model that supports both Indian languages and English, and it said that the work on it has already begun.
Earlier this month, BharatGen, the government backed AI initiative, released a 2.9 billion parameter bilingual LLM, called Param 1. The newly launched LLM, dubbed ‘BharatGen Param 1 Indic Scale’, is a pre-trained base model built entirely from scratch and features a staggering 25% Indic data.
