As the government-backed BharatGen prepares to showcase its next generation of large language models at the India AI Impact Summit, questions around what “sovereign AI” means in practice — and what India is actually building — are becoming central to the policy and technology debate. BharatGen, a consortium of leading academic institutions funded by the government, is working on foundational AI models designed to support Indian languages, data, and use cases.
Rishi Bal, chief executive officer of BharatGen, spoke to Khalid Anzar and Harsh Shivam about why the consortium is focused on building India’s native AI models, how it is being funded, what has been released so far, and what the next phase — including the 17-billion-parameter PARAM 2 model — is meant to achieve. Edited excerpts:
What is BharatGen?
BharatGen is a non-profit foundational AI startup. We are entirely government-funded and are a consortium of nearly 10 academic institutes, including IITs from Bombay, Madras, Kanpur, Kharagpur, Mandi, Hyderabad, IIIT Delhi, IIIT Hyderabad, IIM Indore, and more. We have come together to create this new generation of foundational AI in India.
When you say “foundational AI”, what does that mean in practice?
On paper, AI has been around for decades. I have been working on AI for maybe 25 years in different ways. But it is really only since 2018 or 2019 that we saw a major shift in what AI can do. That shift came from transformer architectures, large amounts of data, and large-scale compute coming together.
That combination created a new kind of capability. We are building that foundational layer in India. We believe this is going to be transformative. This is not day one of AI — it is hour one. We are just getting started, and it will touch every aspect of life.
When something is this foundational, a country should have its own technology and its own story, so that progress is not gated by whether someone else is motivated to build it for you.
India has often been seen as a consumer of technology rather than a creator. Why does this moment feel different
It depends on how far back you look. Historically, we have been producers of technology, science, and mathematics in India. There was a long period when that changed, because of wars, invasions, and other disruptions.
But there is no reason to assume that being a consumer should be the default state forever. Today, we are the world’s fourth-largest economy and moving towards third. This has been building for some time. We now need to think of ourselves differently — as creators and innovators.
AI is one place where that shift can happen. And it is not just AI. You see similar investments in areas like quantum, semiconductors, and manufacturing. Even in areas people do not always notice, like motorcycles, Indian companies are among the largest producers globally. There is a broader movement under way.
Where does BharatGen fit into this bigger picture?
BharatGen is one part of the puzzle. The AI ecosystem is very large. You need people building applications, exporting applications, tuning models, building models, and working on data ecosystems. You need players across the stack. BharatGen is focused on one part of that stack — the foundational models.
What are the advantages of making a sovereign AI model
I usually explain this across several dimensions.
First is sovereignty as a guarantee of access. It has three parts. One is supply security — regardless of policy disagreements, you should have access to the technology. The second is knowing what has gone into the model, so you are not surprised by its behaviour, especially in sensitive sectors. The third is serviceability. If you did not build the model, you do not really have the ability to fix or adapt it when something goes wrong.
If someone releases an open-weights model and you adopt it, do you really control these three things? That is an important question.
The second dimension is what I call “Indianness”. We need support for our languages — all 22 of them and their variations. We need support for our culture and context. Today’s global models largely speak and behave like they come from the west coast of the US. If you ask about a headache, they might tell you to take Tylenol. In India, people think in terms of paracetamol brands like Crocin. These details matter. We should have models that speak like us.
There is nothing wrong with models that reflect other cultures. But we should also have models that reflect ours.
Data availability is often raised as a challenge for India. How do you see that?
Most global data is in English. Only a handful of countries — like China, and maybe Japan — have large volumes of data in their own languages. Everyone else faces this challenge.
Yes, it is difficult. But if we solve this problem for India, we also create methods that can work for many other countries with similar constraints. That means new training techniques, new research, and new ways of working with data.
So where does Indian data come from?
There is Indian data, and we are working on the ground to unlock it. We reach out to publishers, people with old books, and community radio stations. When you think about speech, we are a country of more than a billion people, and we talk a lot. There is a lot of speech data.
The challenge is to unlock it while dealing with privacy concerns, or to build methods that can learn from data in privacy-preserving ways. That is part of what we are doing at BharatGen.
What kind of support is BharatGen getting from the government?
BharatGen received seed funding of ₹235 crore from the Department of Science and Technology. That covered hiring, operations, office space, and other expenses, as well as some compute.
We have now also received a first round of support from the Ministry of Electronics and Information Technology, largely in the form of GPUs. Compute is a major requirement for this kind of work. Overall, BharatGen is entirely government-funded.
Where does model development stand today?
We launched BharatGen on June 2 last year. We released a 2.9-billion-parameter base model called PARAM and made it open source, so any Indian developer can access it.
We also released domain-specific models — for Ayurveda, agriculture, finance, and legal use cases. These include Agri-PARAM, Ayur-PARAM, Finance-PARAM, and Legal-PARAM. PARAM is the base, or “mother”, model, and these are fine-tuned on top of it.
Now we have PARAM 2, which is a 17-billion-parameter model.
Is PARAM 2 already in use?
All the demos you will see at the summit are powered by PARAM 2. We will roll it out after the official launch at the summit.
How does PARAM 2 compare with other models?
It is a 17-billion-parameter model, so it sits in the same range as other 17 to 20 billion parameter models globally. We will publish not just the model, but also the benchmarks, so people can see how it performs.
But most benchmarks are designed around Western datasets, right?
Exactly. When we launched Ayur-PARAM, the first thing we realised was that there were no good benchmarks. So we built one.
The same thing happened with our document-focused vision-language model. Initially, it looked like it was outperforming everything. But our team said the benchmark was too weak and built a tougher one. On that, our model came second instead of first.
That is the culture you need. Everyone wants overnight success, but the reality is that every “overnight success” is usually a decade in the making. If you look at China, it didn't have a DeepSeek moment overnight. That overnight moment also was a decade in the making.
What looks like a sudden breakthrough is the result of years of sustained work and research output. We are on a similar journey.
Are there any consumer-facing applications yet?
We are at the proof-of-concept and pilot stage across areas like governance, education, the Ministry of Culture, finance, and the private sector. Some of these are customer-facing pilots.
People often ask whether this means we are building the Indian ChatGPT. But ChatGPT is not just one model. It is a collection of systems — models, inference infrastructure, data pipelines. You have to build these capabilities step by step.
Where can developers access BharatGen models today?
We are deploying them on AI Kosh and Hugging Face. Developers can access them from either place. We have seen over a thousand downloads already, which suggests people are experimenting with them.
Over time, we will also provide broader inference access. You will also see BharatGen models hosted on partner platforms. For example, at the summit, you will see demos built using our models at cloud provider booths. The goal is to maximise distribution and availability.
Anything to summarise this conversation?
I’ll just say that BharatGen is doing amazing things in making Indian models in Indian languages available to all Indians, and in a way that will be truly accessible. This is just the start of the journey and there is a lot more to come.
 "China's DeepSeek moment to India's sovereign AI, what BharatGen CEO said")
