Technical reports document our own models and benchmarks in detail: the data, the architecture, the training recipe and how we evaluated them. For new methods and findings, see our research papers.
A reference architecture for a sovereign AI stack, written with IBM and NxtGen, and a working voice assistant for farmers built on it, in Telugu, Hindi and English. Shrutam and Sooktam carry the voice, and Param2 was evaluated for reasoning.
Ganesh Ramakrishnan (BharatGen), Amith Singhee (IBM), Abhisyant Anasapurapu (NxtGen) and 17 others
A 2.9 billion parameter Hindi and English model fine-tuned from Param1 on a curated Ayurveda dataset of classical texts and clinical guidance, evaluated on BhashaBench-Ayur.
Sravan Gorugantu, Mohd Nauman, Vijay Devane, Shyam Pawar and 6 others
Presents an AI assistant for Indian farmers that tracks intent across a multi-turn conversation and uses retrieval-augmented generation in Hindi and English.
Abhay Vijayvargia, Ajay Nagpal, Kundeshwar Pundalik, Atharva Savarkar and 4 others
A 2.9B parameter bilingual language model for Hindi and English, with morphology-aware tokenisation and new evaluation benchmarks for Indic code-switching.
Kundeshwar Pundalik, Piyush Sawarkar, Nihar Sahoo, Abhishek Shinde and 18 others
Every model we release comes with a model card on Hugging Face: what it is for, the languages it covers, how it was evaluated and how to run it.
Our flagship text model: a mixture-of-experts model that reasons step by step, across Indian languages.
Model card on Hugging Face ↗VisionOur vision language model for reading and answering questions about documents.
Model card on Hugging Face ↗Speech-to-textReal-time transcription in 12 Indian languages, including speakers who switch languages.
Model card on Hugging Face ↗Text-to-speechNatural, human-sounding speech, with zero-shot voice cloning.
Model card on Hugging Face ↗Peer-reviewed papers and preprints from BharatGen and the consortium.