'Most real-world use cases don’t need trillion-parameter AI': BharatGen CEO Rishi Bal on Param 2 launch

BharatGen on February 16 launched Param 2, a 17-billion-parameter artificial intelligence model designed to serve India’s linguistic and sectoral needs, positioning it as a domestic alternative to global systems from companies such as OpenAI and Google.

The model, unveiled at the India AI Impact Summit 2026, supports 22 Indian languages.

“Param 2 launches today, that’s our 17-billion-parameter mixture-of-experts model, 22 Indian languages, support for code, maths and reasoning,” BharatGen Chief Executive Rishi Bal told Business Today on the sidelines of the summit. “It’s a pre-trained checkpoint… developers can take it and fine-tune it… because Indian languages are natively supported.” 

The government-backed consortium is betting that smaller, specialised models can deliver stronger real-world impact in emerging markets than massive general-purpose systems.

“A majority of business, government and many consumer cases actually don’t need a trillion-parameter frontier model,” Bal said. A trillion-parameter system is “like a thousand PhDs all bottled into one,” he added, but most organisations instead need domain-specific expertise tailored to local conditions. 

India-first architecture

Unlike many global models trained primarily on English-language data, Param 2 was built from the ground up for multilingual Indian usage.

BharatGen plans a rolling release strategy, with additional checkpoints and fine-tuned versions to follow. The model will also serve as a base layer for voice and vision systems.

“We have chosen to build single-modality multilingual models,” Bal said. Param 2 will be “very text-focused” initially, with separate releases planned for voice and document-understanding capabilities. 

Heavy compute investment

Training Param 2 required substantial government-backed computing resources. BharatGen received one of the largest allocations under India AI Mission, including thousands of graphics processing units.

“The GPUs that we’ve used so far from the AI Mission have gone into training of Param 2,” Bal said, noting that the project has been using roughly 4,000 GPUs for over a month. 

The company plans to scale further, with larger models already in development.

“You should expect to see multiple size models… some small ones for edge inference… bigger ones like the 17B… and progressively larger models over the course of this year,” he said.