ARTICLE AD BOX
New Delhi: BharatGen has officially launched its 17-billion-parameter multilingual AI model, which is called Param2, at the India AI Impact Summit 2026, which is set to kick off on February 16 in New Delhi. This model will support 22 Indian languages and is being positioned as part of India’s broader push to develop sovereign AI systems trained on domestic data and run on local infrastructure. This latest model is being developed under the BharatGen consortium, which operates out of the Technology Innovation Hub at IIT Bombay, and it is supported by the Department of Science and Technology.
Param2 is described as a 17-billion-parameter mixture of experts model, which is built to work across 22 Indian languages. It has been trained on the Indian language data and is intended to be used across text, speech, and vision tasks, rather than only as a text-based language model. BharatGen also stated that the model will be demonstrated at the summit via sector-focused apps developed with the government and industry partners. BharatGen has also added that the underlying models are being developed to support reasoning, maths, and code-related tasks, and are being trained using the infrastructure made available under the IndiaAI Mission, along with the data drawn from Bharat Data Sagar repository.
The aim of a large, multilingual model is to reflect a broader policy direction that has been more visible across the multiple government-backed AI projects over the past few years. Public services in India operate across dozens of languages, and much of the information citizens interact with comes in a mix of formats, including text, scanned documents, and speech. Multilingual models are meant to reduce that gap by enabling the same system to work across languages, rather than relying on separate tools or manual translation layers.
BharatGen’s Param2 models sit along with the growing set of India-focused AI efforts that are being framed around the domestic data, local infrastructure, and language coverage. The programme has already released models for text, speech recognition, text-to-speech, and document understanding, and Param2 extends this into the larger, general-purpose multilingual system.
English (US) ·