IIT Bombay-led BharatGPT consortium, in partnership with Vizzhy Inc, has unveiled ‘Hanooman’, a series of Indic large language models trained on 22 Indian languages.“We have built an enterprise-ready family of foundational language models ranging up to 40 billion parameters in terms of size,” Vishnu Vardhan, cofounder of Vizzhy, told ET. “The first four among the series which are 1.5 billion, 7 billion, 13 billion and 40 billion parameters will be released next month and will be open-sourced.”

The models can currently respond in 11 Indian languages including Hindi, Tamil, Telugu, Malayalam and Marathi, he said, adding: “We aspire to extend the model’s capabilities to all 22 Indian languages.”

Vardhan was speaking on the sidelines of the Nasscom Technology and Leadership Forum in Mumbai on Monday, where the announcement was made.

BharatGPT is a research consortium led by IIT Bombay with seven other IITs. It is backed by the Department of Science and Technology and Reliance Jio.

“The engagement with Reliance Jio consists of their industry-specific downstream applications in areas such as telecom and retail and requires the building of smaller customised ML models…,” professor Ganesh Ramakrishnan of IIT-B said.

Hanooman shall have multimodal AI capabilities for generating text-to-text, text-to-speech, text-to-video and vice versa content. Vizzhy is already in talks with BFSI enterprises, healthcare organisations and mobile app providers to offer model-as-a-service or create specialised models by fine-tuning the Hanooman series, Vardhan said.

The first among these fine-tuned versions is healthcare model VizzhyGPT, trained on hospital data of two large hospital chains in India, ET had reported last week.

Among the major challenges for building Indian LLMs is the sourcing of quality datasets in Indian languages, he said.

“The first hurdle for us is to improve the quality of these datasets in Indian languages, whether it is in the form of text, audio or video. Organisations are currently using synthetic datasets which are created by translating from other languages. This may increase chances of inaccuracies or hallucination,” Vardhan said.

Hanooman joins the Indic AI race with other language models such as Ola’s Krutrim, SaravamAI’s OpenHathi and IIT-Madras’s Airavata model.