Sarvam AI launches 30B and 105B models tailored for India-focused deployment

Sarvam AI, one of the select startups chosen under the India AI Mission in 2025 to develop sovereign language models, has released two new models — 30B and 105B. The company said both models are built to operate at scale and have been trained more efficiently compared to other large language models (LLMs). At the India AI Impact Summit 2026, Sarvam AI team members demonstrated the capabilities of both models live.

30B model built on mixture-of-experts architecture

The 30B model is a mixture-of-experts architecture with one billion activated parameters and has been pre-trained from scratch on 16 trillion tokens of text. “Even in the mixture of experts world, we have gone towards systems, which can be very, very efficient for inferencing,” said Dr. Pratyush Kumar, Co-founder, Sarvam AI, while presenting the model. Demonstrating its use, team Sarvam showed how the conversational model-Vikram, named after Dr. Vikram Sarabhai, could answer in Indic languages, effortlessly switching languages between and not losing context, and the model can run on a feature phone. “We think of the 30-billion parameter model as the efficient work horse that will run at scale to power a billion conversations a day and we are looking forward to deploying that across all the things that we do,” Kumar said.     

105B model targets advanced reasoning tasks

The 105B model, designed for more complex reasoning tasks, is also built on a mixture-of-experts architecture. It uses nine billion activated parameters and is estimated to be roughly twice as expensive as the 30B model. “It has a larger context window of 128000 tokens, therefore allows you to do more thinking and do more complex tasks. It is on a par with most other open models and with several closed sources frontier models of its class,” Kumar said during the presentation. Sarvam shared benchmark comparisons of its reasoning performance against other models, stating that on MMLU-Pro — an advanced AI evaluation benchmark — its model outperformed GPT-120B.  

“We find that we are almost close to the best models out there for the model of this size, enabling this to be used also for software engineering where you can fix smaller bugs, write smaller pieces of code,” Kumar added. Sarvam said it will open-source both the new models.