In its mission to build open source LLMs for Indian researchers and developers, BharatGen, the government backed AI initiative, has released a 2.9 billion parameter bilingual LLM, called Param 1.
The newly launched LLM, dubbed ‘BharatGen Param 1 Indic Scale’, is a pre-trained base model built entirely from scratch and features a staggering 25% Indic data—a stark contrast to the mere 0.01% Indic data typically used in models like Meta’s Llama.
You can check out the model on AIKosha.
“Pre-training is an enormous undertaking and often an insurmountable barrier for many. That’s why we’ve taken on this challenge—to provide a robust foundation that you can easily fine-tune for your specific applications,” BharatGen said in a statement.
Developers can now fine-tune the model via AIKosha to build diverse applications ranging from Indic chatbots to India-specific copilots and knowledge systems. “With our 2.9 billion parameter base model, we are unlocking new possibilities for innovation and growth across the nation. We hope this sovereign LLM model checkpoint serves as a foundation for India-specific solutions, enabling developers to fine-tune and shape the next generation of AI applications for Bharat,” Prof Ganesh Ramakrishnan, head of BharatGen, told AIM.
Alongside the LLM, the team also launched 20 new speech models (across 19 Indian language variations)—targeting voice-first interfaces and speech-based innovation for Indian users on AIKosha—the AI innovation repository by MeitY, Government of India.
This includes 9 models under A2TTS-v0.5: Speaker Adaptive TTS. These allow developers to generate speech that matches a provided speaker’s voice, available in Marathi, Bengali, Hindi, Gujarati, Tamil, Kannada, Punjabi, Telugu, and Malayalam.
There are five models under Speaker-Conditioned TTS (pflow) which offer high-fidelity text-to-speech models for Marathi, Tamil, Hindi, Telugu, and Bengali.
Then there are other five models under Voicebox TTS Models, which are adaptable voice synthesis for Hindi, Tamil, Marathi, Telugu, and Bengali.
BharatGen says that these models were built from the ground up with data collected directly for five Indian languages, addressing a major gap in high-quality, publicly available speech models for Indic languages.
AIKosha, launched by Union Minister Ashwini Vaishnaw, is India’s official AI repository and the new home for these models. The repository aims to centralise India’s AI assets and fuel collaborative innovation.
“These foundational models are engineered to supercharge India’s AI research and innovation ecosystem,” BharatGen noted, inviting the community to “build an AI that genuinely speaks to, and for, India.”
Along with Ramakrishnan, the contributors of the model include Kundeshwar Pundalik, Durga S, Prateek Chanda, Vedant Goswami, Atul Kumar Singh, Saral Sureka, Panditi Bhagawan, Ajay Nagpal, Smita Gautam, Pankaj Singh, Rishi Bal, and Prof Rohit Saluja.
While earlier speaking with AIM, Ramakrishnan said that unlike private entities, BharatGen operates with a clear mission—‘GenAI for Bharat, by Bharat’. With an investment of under ₹235 crores, which is close to $27 million, leveraging cost-efficient computing and attracting top talent from graduates from IITs.
Read: This Govt-Funded ₹235 Cr AI Initiative is India’s Real Answer to DeepSeek
The BharatGen consortium comprises IIT Bombay, IIT Kanpur, IIT Mandi, IIT Madras, IIT Hyderabad, IIIT Hyderabad, and IIM Indore.
Vaishnaw had earlier stated that India would have its own foundational AI models within 7-8 months, and BharatGen is a key part of that vision. “Yes, we are very much on track. The Minister has been briefed, and we are aligned with the timeline,” Ramakrishnan had confirmed earlier. Param-1 is a clear sign of that roadmap.
“Our goal is not just to build AI models but to provide resources that startups and system integrators can leverage,” said Ramakrishnan.
Last month, MeitY also selected Sarvam AI under the IndiaAI Mission to develop India’s sovereign LLM as part of the effort to create indigenous AI capabilities. The team had proposed the development of a 70-billion parameter multimodal AI model that supports both Indian languages and English, and work on it has already begun.
