When it comes to building sovereign AI models in India, Sarvam AI has emerged as one of the key players. Founded in July 2023 by Vivek Raghavan and Pratyush Kumar, the Bengaluru-based startup aims to make generative AI accessible at scale across India. Unlike many startups that fine-tune existing global models, Sarvam is building its own foundation models from scratch.
In April 2025, the company was selected under the IndiaAI Mission to help develop India’s first indigenous large language model. Sarvam’s goal is to create an AI ecosystem that is built in India, trained on Indian languages and data, and deployed to serve Indian needs.
The start-up also showcased AI-powered glasses, which Prime Minister Narendra Modi tried at the event. We spoke with Vivek Raghavan about Sarvam’s challenges and how its models compare with global platforms like ChatGPT and Gemini. In an interview with Rakesh Kumar, Vivek Raghavan shared the challenges Sarvam faced in building these models and discussed how they compare with global platforms such as ChatGPT and Gemini.
What were the most significant technical and operational challenges you faced while building a foundation model of this scale in India?
Building a foundation model is an extremely complex task. Our team, led by Pratyush and supported by many talented young developers, has worked very hard to make this possible. It is not an easy undertaking.
We trained the model on a very large dataset of nearly 18 trillion tokens. The models range in size from 30 billion to 105 billion parameters. Training systems of this scale is highly challenging, especially since this is the first time such an effort has been undertaken in India.
More importantly, this model serves as a foundation. We hope it inspires many more teams in the country to build similar systems. This is something India must do, and while Sarvam is taking the lead, many others will follow.
Were there funding issues?
Under the IndiaAI Mission, GPUs were made available to us to train the model, which was helpful. However, the challenges go beyond funding.
Whenever something is being done for the first time in the country, there are multiple technical and operational hurdles. It is important for India that such successes are widely recognized, because building a foundation model of this scale is not something that just anyone can accomplish. It requires deep expertise, sustained effort, and strong teams.
How do Sarvam’s models compare with international models?
Globally, there are very large models, particularly from countries like the US and others, which are significantly bigger than ours and backed by investments running into hundreds of billions of dollars. Naturally, such models may demonstrate certain additional capabilities.
However, within the 100-billion-parameter category, we believe our model can be compared with some of the best in the world.
Which industries or sectors are currently adopting Sarvam AI’s models, and what kind of real-world impact are you seeing?
Yes, Sarvam AI is already being used across multiple sectors. For example, our voice AI systems handle more than one million minutes of conversations every day.
Enterprise adoption has been particularly strong, with many organizations integrating our AI solutions into their operations.
Why is it strategically important for India to develop its own sovereign AI models rather than rely solely on global platforms?
AI is a strategic technology. The simplest way to explain it is that AI today is like nuclear power. There are countries that possess nuclear capability and those that do not. AI will create a similar divide.
India must decide which side of that divide it wants to be on.
Can we expect this to scale up?
Yes, absolutely. The real challenge is not just building the model, but scaling it for widespread use. Creating a model is one step; ensuring that it reaches people and is used effectively is far more important. We now need to focus on how to put this technology into the hands of users at scale, while also managing the associated costs.