BharatGen Advances India’s AI Ambitions with Multilingual Models and Data Initiatives

By: Harsh Vardhan

|

Updated on: 06-Jul-2026 02:00 PM

6,496 views

Follow Us:

6,496 views

BharatGen is leading India’s efforts to develop sovereign foundational AI models, responding to global skepticism about the country’s capabilities. The initiative gained momentum after OpenAI CEO Sam Altman’s 2023 remarks in New Delhi, where he suggested it was “totally hopeless” for an Indian startup to rival ChatGPT. BharatGen’s team, initially a small group, has grown to over 60 members and now receives approximately Rs 1,235 crore in state funding through the Department of Science and Technology (DST) and the IndiaAI Mission.

Key Highlights

• BharatGen receives over Rs 1,235 crore in state funding for AI development.
• 17-billion parameter model uses Mixture of Experts for efficient multilingual processing.
• Bharat Data Sagar collects over 20 trillion tokens of regional data.
• Patram-Bench evaluates AI performance across 22 Indian languages.
• BharatGen models deployed in government, healthcare, and education sectors.

BharatGen’s Formation and Growth

Professor Ganesh Ramakrishnan, who chose to pursue his PhD at IIT Bombay, played a key role in building BharatGen’s ecosystem. In early 2023, a consortium-driven approach was proposed by former Telecom Secretary K. Rajaraman, aiming to unite Indian academia and industry. BharatGen was established as a Section 8 not-for-profit consortium, anchored by IIT Bombay and spanning nine leading academic institutions. The team received an initial Rs 235 crore grant from the National Mission on Interdisciplinary Cyber-Physical Systems (NM-ICPS), working without salaries until funding arrived.

BharatGen’s models are designed to deliver stable AI performance at 2.9-, 5-, and 7-billion parameter thresholds. A significant funding boost of nearly Rs 1,000 crore from the IndiaAI Mission has supported the development of advanced AI infrastructure and talent.

Technical Innovations and Language Challenges

BharatGen addressed the limitations of existing tokenisation methods, which are primarily designed for English and often inefficient for Indian languages. The team developed phonetically aware tokenisation layers to improve model efficiency and reduce computational costs. Their 17-billion parameter Param 2 text model uses a Mixture of Experts (MoE) framework with 64 experts, including two shared experts for code-mixing and domain-mixing. This approach activates only 2.4 billion parameters per token transaction, lowering inference costs.

The Shrutam speech model, India’s first LLM-based Automatic Speech Recognition (ASR) system, can handle regional accents and duration prediction. During pre-training, the MoE layers naturally grouped related languages, such as Hindi and Marathi, improving cross-lingual performance. Shrutam also addresses challenges in text-to-speech by capturing pronunciation variations across regions.

Data Acquisition and Evaluation

Much of India’s domain knowledge exists outside the digital realm. BharatGen launched Bharat Data Sagar, a street-level data acquisition network, to collect regional texts and manuscripts. The initiative has gathered over 20 trillion tokens, approaching 3 petabytes of data. To ensure data quality, BharatGen developed automated pipelines for metadata verification and de-duplication.

The Patram-Bench evaluation framework includes 52,000 documents, 2.4 billion question-answer pairs, and 52 task types across 22 Indian languages. Every token’s provenance is tracked for compliance, especially under the Digital Personal Data Protection (DPDP) Act.

Deployment and Future Plans

BharatGen’s models are already deployed in public and private sectors. They power MahaGPT for the Maharashtra government, Gyan Bharat for the Ministry of Culture, and support regulatory, healthcare, and educational applications. BharatGen is not focusing on consumer-first smartphone apps but aims to make AI cost-effective, targeting operational costs as low as 10 Paisa per minute for rural users.

The team is developing a trillion-parameter model to serve as a knowledge repository, with smaller models distilled for specific tasks. BharatGen has joined Project Tapestry to promote open, accessible AI and collaborates internationally to ensure AI sovereignty. The initiative emphasizes frugality and local relevance, aiming to make India a foundational innovator rather than just a user of foreign technology.