India's Sovereign AI Goes Public
On June 15, 2026, at Bharat Innovates 2026 in Nice, France, IIT Bombay formally presented BharatGen to the world — a sovereign AI ecosystem built for India's 1.4 billion people across all 22 scheduled languages.
The announcement, which drew 118,900 views on X within hours, represents the culmination of a multi-year national effort: 9 premier academic institutions, 60+ researchers, engineers, and linguists, backed by India's Department of Science and Technology, the IndiaAI Mission, and ₹988.6 crore in funding.
BharatGen is not a single model. It is a four-family ecosystem covering the full stack of language interaction — text, speech, and documents — in every officially scheduled Indian language.
The Four Model Families
Param2 — Foundational Text Model
The cornerstone of BharatGen. Param2 is a foundational large language model that works across all 22 scheduled Indian languages with:
- Reasoning capabilities (multi-step problem solving)
- Coding support
- Tool calling for agentic use cases
Param2 is built to handle not just translation but genuine native-language understanding — the cultural nuances, idioms, and domain knowledge that Western models trained primarily on English-language data routinely miss in Indian language contexts.
The use cases highlighted span governance, healthcare, education, insurance, finance, and cultural preservation — domains where the language gap between frontier AI and India's actual population has been most acute.
Shrutam2 — Multilingual Speech Recognition
Automatic speech recognition across Indian languages. India is predominantly an oral culture in many regions — literacy rates vary significantly, and for hundreds of millions of users, voice is the primary interaction modality.
Shrutam2 addresses the specific challenges of Indian speech recognition: phonetic complexity, tonal variations, code-switching (mixing multiple languages within a single utterance), and the acoustic diversity across India's geographic spread.
Sooktam2 — Text-to-Speech with Voice Cloning
Text-to-speech synthesis across Indian languages, with a notable capability: zero-shot voice cloning. The model can reproduce a target speaker's voice characteristics without fine-tuning on that speaker's data — enabling personalised speech synthesis for applications ranging from accessibility tools to personalised education.
Zero-shot voice cloning in multilingual Indian language contexts is technically demanding — Indian language prosody and phonology differ substantially from the Latin-script languages where most voice cloning research has been conducted. This is a meaningful technical achievement.