India AI Impact Summit: Sarvam calls for self-dependence in LLMs

Vivek Raghavan says India must balance building larger language models with driving real-world adoption, as usage and applications will determine the success of a sovereign AI stack

Sarvam AI’s recently released language models have drawn praise from Google Chief Executive Officer Sundar Pichai. Sarvam Cofounder Vivek Raghavan, in an interaction with Avik Das, talks about their effectiveness, why the real challenge lies in deployment, and India’s push to build a sovereign artificial intelligence (AI) stack. Edited excerpts:

 

How has the Indian AI ecosystem and the government reacted to the language models that you launched? 

I think everybody is excited that India has been able to build a large language model (LLM) like this from scratch, with no data dependency on anyone else — something that is strong in Indian languages and comparable to other models of its size. I am not saying it is comparable to the largest Gemini or Anthropic models. What we tried to do was make it a very efficient model. It can be the harbinger of bigger things to come.

 

Would you want to build bigger models going ahead? 

We do want to, but what is more important is that these models are actually used, and that we figure out how. Especially in Indian languages, and for about 95 per cent of the tasks people need from an LLM, these models are perfectly suited. The smaller model (30 billion parameters) is very good at handling Indian conversations at a very low cost. The larger model (105 billion parameters) is a reasoning model that is strong in coding and can be used for more complex web searches.

 

Should the country’s focus be on large models or applications? 

I don’t think it’s an either/or situation. You have to work on applications because that’s where you demonstrate the value of the models. There’s no doubt about that. But if you abdicate your ability to build models, you put yourself in a position of dependence on others — and they will dictate terms.

 

How do we make sure these models are used? Have you had any conversations with the government? 

I’m sure there are many discussions, but nothing that I can disclose at this time.

 

Is there scope to expand these models? Should the country look to build larger models or AI applications? 

If the opportunity arises, we certainly could look at that. The bigger question is whether India should build models larger than these. Of course it should. There are many applications where a small model can be very effective, but there are also areas where a large model is extremely useful for the country. I don’t know whether I can make a business case for doing that from scratch, but as a country, I think there is a case to be made.

 

Since building larger models is more expensive, would you need to raise funding for that? 

That is certainly something we have looked at. But our primary focus is to leverage the models we already have and serve them at scale. We will measure ourselves by how widely our models are used, and that will be one of our key challenges.