BharatGen turns two this month. On 14 October, the team and colleagues from across the consortium gathered at our office at IIT Bombay to mark it. Each team stood up in turn and talked through what it had built, what it had learnt, and what it wants to do next. There was cake.
It was a good moment to look back. BharatGen launched in New Delhi in October 2024 with a plain goal: to build AI models for India from the ground up. This post tells the story of the two years since, phase by phase.
Two years at a glance
- Four families of open-weight AI models, across text, speech and vision, that anyone can download, adapt and run.
- All 22 scheduled languages of India covered by Param2, our flagship language model, along with English.
- Nine of India's top academic institutions in the consortium, anchored at IIT Bombay.
- More than 60 full-time engineers, researchers, linguists and builders, and more than 120 interns trained.
- Partners at home and abroad: hospitals, state governments, a ministry and a newsroom in India, and research partners in Japan and Australia.
Starting out
Dr Jitendra Singh, Union Minister of State for Science and Technology, launched BharatGen in New Delhi in October 2024, with Prof. Abhay Karandikar, Secretary of the Department of Science and Technology, and the consortium institutions. Backed by the Department of Science and Technology, the consortium brought together IIT Bombay, IIIT Hyderabad, IIT Mandi, IIT Kanpur, IIT Hyderabad, IIM Indore, IIIT Delhi, IIT Madras and IIT Kharagpur.
The starting question was simple to state and hard to answer. Global AI models are largely trained on English and a handful of other widely used languages. They tend to break words in Indian languages into many small pieces, and they often miss Indian context altogether. Building AI models for India meant starting from scratch: gathering data, designing tokenisers for Indian scripts, and training AI models from the ground up rather than adapting someone else's.
The launch is on record, in this video from the Department of Science and Technology.
First applications
Even before the first AI models were released, the consortium was showing what they could do for real people. In May 2025 two applications went up on our website.
- Krishi Sathi, a WhatsApp assistant that farmers talk to in voice notes, in any of India's 22 scheduled languages, for advice on pests, crops and storage.
- e-VikrAI, which turns a single photo of a product into a complete online listing in Indian languages, cutting the time to list a product from about six minutes to about 15 seconds.
Both taught lessons that stayed with us. Krishi Sathi's first users asked questions nobody had planned for. e-VikrAI showed that the hardest part of selling online is often paperwork and language, not the product.
First AI models
The first public results came at the BharatGen Summit, held on 2–3 June 2025 at the Dr Ambedkar International Centre in New Delhi. Three AI models launched there:
- Patram 1, our first vision language model (VLM).
- Shrutam 1, for Hindi speech-to-text.
- Sooktam 1, for text-to-speech in nine Indian languages.
The summit also hosted the final of the BharatGen Generative AI Hackathon, which drew 279 student teams from 21 states, an early sign of how many people wanted to build with AI models made for India.
A month later, in July 2025, came Param1, our first foundation AI model, trained from scratch on 7.5 trillion tokens of Hindi and English.
National backing, and a family of AI models
In September 2025 the work was given national scale. The IndiaAI Mission of the Ministry of Electronics and Information Technology backed BharatGen, announced by Union Minister Ashwini Vaishnaw in New Delhi. The same week, IBM and BharatGen announced a collaboration on Indic AI solutions, data pipelines and benchmarks.
The months that followed were about breadth. In August 2025, Patram 1.1 added multi-turn conversations about an image and the ability to pull out key fields such as names and amounts. From August to October 2025 came a set of domain models: AyurParam, AgriParam, FinanceParam and LegalParam. Each arrived with a benchmark from BhashaBench, so that anyone can test AI models on Indian knowledge in that field, not only on general questions.
On 7 November 2025 the BharatGen Technology Foundation was incorporated as a Section 8 not-for-profit company, giving the work a lasting home. In December 2025 the Param1 family grew: Param1 5B for Hindi and English, and Param1 7B, a mixture-of-experts model covering English and 15 Indian languages.
The second generation
In January 2026, Prime Minister Narendra Modi met the leaders of 12 Indian organisations building foundation AI models, BharatGen among them, at a roundtable in New Delhi ahead of the India AI Impact Summit.
The summit itself, at Bharat Mandapam in February 2026, was the biggest moment of the year. The Prime Minister stopped at the BharatGen pavilion, and our second generation of AI models arrived:
- Param2 covers all 22 scheduled Indian languages and English. It is a mixture-of-experts model with 17 billion parameters in total and about 2.4 billion active for each token, built from scratch with its own tokeniser and released with open weights on Hugging Face and AIKosh.
- Shrutam 2 understands 12 Indian languages, and Sooktam 2 speaks 11 plus Indian English.
- Patram 1.2 carried the vision work forward.
Partners used the summit to show what they were building. The India Today Group unveiled Sutra, an AI news anchor built with BharatGen's AI models. The Maharashtra Institution for Transformation (MITRA), with IIT Bombay and BharatGen, launched an assistant that searches more than 150,000 government resolutions and regulations, now in beta as MahaSamvaad. The Government of Andhra Pradesh signed an agreement with BharatGen to set up a Telugu-first AI tech hub. And a panel on Gyan Bharatam discussed how AI can help bring India's manuscripts back into use.
Into hospitals, ministries and data centres
After the summit, the focus turned to putting AI models to work.
- At Amrita hospitals, Amrita Healthcare's MedSum, built on BharatGen's open-weight AI models, transcribes consultations and drafts clinical notes, so that doctors can give patients their full attention.
- With the Ministry of Culture, an active pilot for Gyan Bharatam is building AI tools to read manuscripts, answer questions about them and help students pronounce Sanskrit correctly.
- In March 2026, BharatGen signed an agreement with L&T Semiconductor Technologies and L&T-Vyoma to design, build and deploy a sovereign AI compute platform for India, bringing Indian chips, data centres and AI models together.
The pattern is the one we hoped for: organisations that know a field from the inside, building on open-weight AI models made in India.
Going global
In 2026 the work also reached beyond India. In May, BharatGen helped shape Project Tapestry at its first workshop in Paris, and in June it joined as a founding contributor. Project Tapestry helps nations pool data, compute and talent while each keeps control of its own AI models and data.
June also brought two three-way agreements between IIT Bombay, BharatGen and partners abroad: with Japan's National Institute of Informatics (NII), on language models with stronger scientific reasoning, and with Australia's Monash University, on trusted infrastructure and culturally inclusive AI. The same month, our vision team presented research at CVPR, with further papers accepted at ECCV and ICDAR.
On 1 September 2026 came Tapestry's Milestone Zero. BharatGen led the first proof of concept with Monash University, training one model across India and Australia without moving any training data between them. It showed, in practice, an idea that matters to many countries: working together on AI without giving up control of data.
The people
None of this happens without people. BharatGen today has more than 60 full-time engineers, researchers, linguists and builders, and has trained more than 120 interns, from undergraduates to postdoctoral researchers. Linguists work alongside engineers because building for 22 languages is a language problem as much as a computing one.
"It's not just about building the models. It's not just about building the applications. It's about building the builders."
Rishi Bal, CEO, BharatGen
Behind the team stands the consortium. Its faculty, students and research groups have shaped the work from the first day, from the vision team at IIIT Hyderabad to the on-the-ground data collection led from IIM Indore. More on the team and the consortium is on our about page.
What two years taught us
At the town hall on 14 October, the same few lessons came up again and again.
- The field teaches what the lab cannot. Krishi Sathi was trained to help farmers grow crops. One of its first users wanted help storing them. Real users ask questions no benchmark predicts, so we now test with real people early, in deep pilots with one or two partners, before thinking about scale.
- Measure what matters for India. Standard benchmarks rarely test Indian law, agriculture, finance or Ayurveda. So we built BhashaBench to test AI models, including our own, on the knowledge Indians actually need.
- Partners build the best applications. Doctors at Amrita, editors at India Today and officials in Maharashtra know their work far better than we do. Our job is to give them AI models good enough, and open enough, to build on.
- Open weights earn trust. Hospitals and governments need to run AI on their own terms, with their own data. Open-weight AI models make that possible.
What comes next
Two years in, the reasons are the same as on the first day. India needs AI models built for India: for its languages, and also for its documents, its voices, its fields of knowledge and its ways of working. Those AI models need to be built from scratch, so that the choices about data, tokenisers and training are made with India in mind. And they need to be built in the open, so that hospitals, newsrooms, ministries, researchers and students can take them, adapt them and run them on their own terms.
The year ahead is about depth: stronger AI models in more Indian languages across text, speech and vision, and more partners taking them from pilot to everyday use.
Thank you to everyone who has been part of the first two years: the team, the interns, the consortium, our backers at the Department of Science and Technology and the IndiaAI Mission, and every partner who has built something with us. The work goes on, one model and one language at a time.
