The global race for artificial intelligence often feels like a closed-circuit marathon dominated by a few giants in the West and China, but Rishi Bal, CEO and VP of BharatGen, is ready to introduce a new contender. Ahead of the India AI Impact Summit, Bal describes India as a “sleeper” in the AI space – a nation with untapped talent and a unique mission for “digital self-determination”. Operating as a government-funded non-profit, BharatGen isn’t just building another chatbot; it is constructing a sovereign foundation designed to ensure that India’s digital future cannot be “turned off” by shifting global policies or foreign dependencies.
“BharatGen is India’s sovereign foundation model builder, but it’s also a lot more than that. BharatGen’s mission is to raise the overall AI ecosystem in India. And that’s part of the reason that we constituted this company as a non-profit. BharatGen is entirely government funded.”
Also read: AI Impact Summit 2026: PM Modi invites global data to “reside in India”
The crown jewel of this effort is the Param 2 17B model, an architecture built “from first byte to final model” entirely within India. Unlike many global models that treat Indian data as a 2% footnote, Param 2 is infused with up to 50% Indian-centric data. Bal is quick to distinguish this from simply fine-tuning an existing model; this is a ground-up engineering feat designed to handle the linguistic nuances of 22 different languages.
To navigate the country’s diversity without dominant languages like Hindi “washing out” dialects like Tulu or Marathi, BharatGen utilizes a Mixture of Experts (MoE) architecture. This system mimics the human brain by activating only the specific “neurons” needed for a task, which Bal notes is essential for making AI cost-effective and sustainable in a price-sensitive market. By avoiding “dense” models where every neuron fires for every question, they reduce the immense compute costs typically associated with frontier AI.
“Our 17B model is inherently a lot more Indian. So we use somewhere between 25 and 50 percent of Indian source data or Indian related data compared to some of the large language models trained abroad, which we estimate between two and five percent. In India, you have to be thinking about cost from day one. And so that MOE architecture gives us the benefit of both superior performance and then also lower cost.”
Solving the “data paucity” of Indian languages required BharatGen to go beyond the surface of the internet. The team has launched a massive ground-level operation, partnering with local radio stations to capture authentic accents and digitizing ancient manuscripts with advanced OCR techniques. This “Bharat Data Saga” ensures that the models are trained on authentic human interactions rather than falling into the trap of low-quality synthetic data that lacks cultural context.
This commitment to factual reliability is further bolstered by Param 2’s status as a “thinking model”. Before providing an answer, the AI undergoes a reasoning process to verify its own logic, a crucial safeguard for high-stakes applications like the “Bharatiya Nyaya Samhita” legal system or agricultural advisory where a hallucination could have real-world consequences. Bal admits that while the hallucination problem isn’t fully solved globally, their post-processing and verification steps significantly mitigate the risk.
“We actually have a team on the ground that is working with publishers, radio stations, media outlets to actually have them contribute data so that we can actually digitize it and be able to use it for LLM building. Take a 17B model, it’s actually a thinking model. And so one of the things that we do before we spit out an answer is that we actually go through a reasoning process.”
The genesis of BharatGen lies at the intersection of academia and industry, utilizing a consortium of premier institutes like IIT Bombay and IIT Madras. This partnership isn’t just about research; it’s about “building the builders”. Over 100 interns are currently learning to develop LLMs while still in college, ensuring India has a pipeline of talent that understands the intersection of deep research and heavy engineering.
As the Prime Minister prepares to launch these models on the global stage, Bal envisions BharatGen as a piece of Digital Public Infrastructure. By releasing models on Hugging Face – including vision models and domain-specific versions for highways and finance – they are providing a sovereign platform for Indian startups to innovate without relying on “black box” technology from abroad. The goal is to move past the era of being a consumer of foreign tech and prove that India’s specific approach to AI, focused on sovereignty, Indianness, and accessibility is a global game-changer.
“I suspect a lot of people just may not have, I think India is a sleeper in the AI space today, right? I don’t think they understand the level of talent and execution that’s happening on this front. This AI summit would be a great opportunity for industries and governments to see the different approach that India has taken to AI, and hopefully to find ways to partner and move the state of the art forward together.”
