An AI company building India's own AI models
We pretrain text, speech and vision models across Indian languages, publish them with open weights, and work with organisations that build on them.
An AI company, built for India
BharatGen Technology Foundation is an AI company, set up as a Section 8 not-for-profit and anchored at IIT Bombay. We work with a consortium of nine of India's top academic institutions, backed by the Department of Science and Technology and the IndiaAI Mission (MeitY).
Our 60+ engineers, researchers, linguists and builders work from Mumbai, Hyderabad, Pune and Delhi, and build every model from scratch:
- Collect dataCurated with provenance, much of it from the ground
- Design tokenisersBuilt for Indian scripts
- PretrainFrom scratch, on our own corpus
- EvaluateOn benchmarks built for India
- ReleaseOpen weights on Hugging Face and AIKosh
India should make AI, not just use it
Built, not borrowed
Our models are pretrained from scratch, not fine-tuned from someone else's. That way India owns the whole stack, from data to weights.
An ecosystem, not just models
Open weights, open benchmarks and a consortium of institutions, so startups, enterprises, governments and researchers can build on what we make.
For every Indian
AI that works in the languages people actually speak, on the documents they actually use, not only for English-speaking cities.
Two years, four families of AI models
From launch in 2024 to open-weight AI models for text, speech and vision, and the partnerships along the way.
- BharatGen launches with DST's backingLaunched in New Delhi on 30 September under the Department of Science and Technology's NM-ICPS mission, with ₹235 crore (about US$27 million at ₹88 = US$1) through the Technology Innovation Hub at IIT Bombay, which leads a consortium of India's top academic institutions.
- First applications at India Mobile CongressWithin our first 100 days, we show e-VikrAI and Krishi Sathi at IMC 2024 in New Delhi, and demonstrate both to the DST Secretary.
- Patram1, Shrutam1 and Sooktam1Our first vision language model, Shrutam1 for Hindi speech-to-text, and Sooktam1 for text-to-speech in nine Indian languages.
- Param1Our first foundation AI model, trained from scratch on Hindi and English.
- Patram1.1Adds multi-turn conversations and pulls out key fields such as names and amounts.
- Domain models and BhashaBenchAyurParam, AgriParam, FinanceParam and LegalParam, each with a benchmark of Indian knowledge.
- A larger Param1 familyParam1-5B for Hindi and English, and Param1-7B, a mixture of experts for English and 15 Indian languages.
- Param2, Shrutam2, Sooktam2 and Patram1.2Our second generation, launched at the India AI Impact Summit. Param2 covers all 22 scheduled languages.
- Project TapestryWe help shape Project Tapestry at its first workshop in Paris, and join as a founding contributor in June.
- Agreements with NII and MonashThree-way agreements with IIT Bombay and Japan's National Institute of Informatics, and with Australia's Monash University.
- Research at CVPR, ECCV and ICDAROur vision team presents at CVPR, with further papers accepted at ECCV and ICDAR.
- Tapestry's Milestone ZeroWith Monash University, we train one model across India and Australia without moving any training data.
- Ensuring AI speaks everyone's languageWe sign the Gates Foundation's joint commitment with organisations worldwide: within five years, 3.4 billion people whose languages AI underserves should be able to use it in their own language and voice.
- Tapestry Milestone OneTraining across the consortium, with work on cultural alignment and data governance. Started in September and expected to wrap up around the end of the year.