The International Institute of Information Technology, Hyderabad (IIIT-H) has developed and launched Patram-7B-Instruct, the country’s first vision-language foundational model, specifically designed for understanding documents. This cutting-edge model is designed to interpret scanned papers and photographed texts using natural language queries, providing a powerful solution for sectors that deal with diverse and complex Indian document formats.
Patram is part of the broader BharatGen initiative, a government-backed effort to create indigenous multimodal AI models, supported by the Department of Science and Technology (DST).
Built in just five months by a dedicated group of IIIT-Hyderabad alumni and student interns, with support from TiH-IoT at IIT Bombay, Patram-7B is a 7-billion parameter model that is already turning heads for its performance. Despite its relatively compact size, the model has shown strong results on international benchmarks such as DocVQA and VisualMRC, and has been tested successfully on Patram-Bench—a custom dataset designed to reflect real-world Indian document scenarios.
“Patram is a bold step toward making India self-reliant in AI. It bridges the gap between language and visual understanding, particularly for Indian documents,” said Prof. P. J. Narayanan, Director of IIIT-Hyderabad.
The model is open-sourced and now accessible via Hugging Face and AIKosh, India’s national AI platform. Its launch marks a key milestone in India’s push for open, indigenous AI infrastructure under initiatives like Digital India and Atmanirbhar Bharat.
Also Read: Rajasthan reimagines student assessments with AI & NEP framework
In addition to Patram, the team also introduced DocBodh, a generative AI toolkit focused on document intelligence for use in governance, education, law, and business.
Dr. Ravi Kiran Sarvadevabhatla, associate professor at IIIT-H and lead researcher, added, “With Patram and DocBodh, we are building AI that understands India—not just linguistically, but structurally and culturally.”
