Notes from the team building India's AI models
Model launches, case studies, benchmarks and research, written by the people who did the work.
What if your AI could read a document the way you do and show you exactly what it found?
M3Grounder, from BharatGen's vision team and IIIT Hyderabad, answers questions about documents and highlights the exact evidence on the page. Presented at CVPR 2026.
Gyan Bharatam: bringing India's manuscripts back into use
In an active pilot with the Ministry of Culture, BharatGen is building three AI tools for Gyan Bharatam: manuscript reading, conversation over texts, and Sanskrit pronunciation feedback.
MedSum at Amrita hospitals: AI that drafts the notes so doctors can listen
Amrita Healthcare built MedSum on BharatGen's open-weight AI models to transcribe consultations, draft clinical notes and prescriptions, and summarise patient documents for doctors.
Sutra: the AI news anchor built with India Today
The India Today Group unveiled Sutra, an AI news anchor built with BharatGen's AI models, at the India AI Impact Summit in New Delhi on 17 February 2026.
Param2: what we built and how
Param2 is a 17-billion-parameter reasoning model built from scratch for India. It uses about 2.4 billion parameters per token and covers all 22 scheduled Indian languages plus English.
Incorporation of BharatGen Technology Foundation
BharatGen Technology Foundation, a Section 8 not-for-profit company, was incorporated on 7 November 2025. Meet its CEO and Board of Directors.
Introducing BhashaBench V1: Testing AI on what really matters to India
BhashaBench V1 is a bilingual benchmark of 74,166 English and Hindi questions across agriculture, law, finance and Ayurveda. Tests of 29 models show wide gaps by domain and language.
FinanceParam: making Indian financial knowledge accessible to all
FinanceParam is a 2.9 billion parameter English-Hindi model fine-tuned from Param1 on 24 million finance samples, covering tax, banking, investments and policy.
BhashaBench‑Legal: measuring AI's command of Indian jurisprudence
BhashaBench-Legal tests AI models on 24,365 English and Hindi questions from 50+ Indian law exams. Top models pass 75% in English, but procedure, emerging areas and Hindi remain weak spots.
LegalParam: making Indian legal knowledge easier to reach
LegalParam is a 2.9 billion parameter English-Hindi model fine-tuned from Param1 on 5 million Indian legal question-answer pairs, and evaluated on BhashaBench-Legal.
BhashaBench‑Ayur: evaluating AI's knowledge of Ayurveda
BhashaBench-Ayur tests AI models on 14,963 English and Hindi Ayurveda questions from 50+ exams. The best models score around 58%, with the largest gaps in Panchakarma and Dravyaguna.
AyurParam: a specialised language model for Ayurvedic knowledge
AyurParam is a 2.9 billion parameter English-Hindi model fine-tuned from Param1 on about 4.8 million Ayurveda samples. It outperforms similar-sized models on BhashaBench-Ayur.
IndiaAI Mission backs BharatGen with ₹1,058 crore
The IndiaAI Mission has allocated ₹1,058 crore (about US$120 million at ₹88 = US$1) to BharatGen to build AI models for India across text, speech and vision.
IBM and BharatGen collaborate to accelerate AI adoption in India, powered by Indic large language models
IBM and BharatGen announced a collaboration to build Indic AI solutions, data pipelines and benchmarks using BharatGen's models and IBM's AI technologies.
BhashaBench‑Finance: testing AI on India's financial knowledge
BhashaBench-Finance tests AI models on 19,433 questions from Indian finance exams in English and Hindi. Even the best models struggle with quantitative problems and score lower in Hindi.
AgriParam: a bilingual language model for Indian agriculture
AgriParam is a 2.9 billion parameter English-Hindi model fine-tuned from Param1 on 12 million agriculture samples, built for farmer advisories, schemes and rural extension work.
BhashaBench‑Krishi (BBK): measuring AI's gap in Indian agricultural knowledge
BhashaBench-Krishi tests AI models on 15,405 English and Hindi questions from 55+ Indian government agricultural exams. Even top models fall well short on regional and policy-aware farming knowledge.
Param1: a 2.9B‑parameter bilingual language model built from the ground up
Param1 is our 2.9 billion parameter Hindi-English language model, trained from scratch on 7.5 trillion tokens. Here is how we built it and how it performs.
Krishi Sathi: farm advice on WhatsApp, by voice
Krishi Sathi is a WhatsApp assistant built on BharatGen's AI models. Farmers ask about pests, crops and storage in a voice note, in their own language, and get practical advice back.
e‑VikrAI: a product listing from a single photo
e-VikrAI turns one photo of a product into a complete online listing, in Indian languages. It runs on IndicVikr-7B-Instruct, a vision language model built for Indian e-commerce.
BharatGen launches, backed by the Department of Science and Technology
BharatGen launched in New Delhi on 30 September 2024 under the Department of Science and Technology's NM-ICPS mission, with IIT Bombay leading a consortium of academic institutions to build AI models for India.More from BharatGen
Press coverage, talks and interviews are on our News & media page.