India launches three homegrown AI models, including two built to work offline

India has launched three homegrown AI models under the IndiaAI Mission, pairing advanced analytical work with assistants designed to operate without an internet connection. Soket Labs’ EKA-1 targets mathematics, programming and reasoning. ZenteiQ AiTech Innovations’ Dhir 2B and Dhir 5B instead emphasize document work and everyday automation, including offline use on compatible phones and laptops.

The three models were jointly introduced at an event in New Delhi with support from the Ministry of Electronics and Information Technology, according to Ascendants. IT Ministry Secretary S. Krishnan highlighted potential uses in government departments, educational institutions, research facilities and private businesses.

EKA-1 takes the analytical route

EKA-1 is the largest model in the launch. Its reported specifications include 26.3 billion parameters and training on approximately 7.5 trillion tokens of data. Its stated language coverage spans English and all 22 scheduled Indian languages. That gives it broader listed language support than the Dhir pair, which covers English and 10 Indian languages.

Developed by Soket Labs, EKA-1 is described as an open, sparse hybrid language model. Its intended work includes mathematics, software programming, logical reasoning and multi-step AI agent operations: sequences of tasks rather than only individual replies. The applications identified for it include banking, financial services and cybersecurity, where the launch positions it for advanced analytical processing.

Dhir brings document work onto the device

ZenteiQ’s Dhir 2B and Dhir 5B are described as open-source models aimed at smaller-scale applications and routine productivity. Their listed jobs are concrete: search through documents, extract relevant information and automate repetitive tasks.

Vajiram & Ravi also describes the Dhir models as deployable on an organization’s own premises, with support for computing platforms including GPUs and TPUs. It identifies document-based automation and citizen or enterprise services across defence, government, industry and education as intended uses, particularly where connectivity is limited or information is sensitive.

Local processing can keep sensitive information on the device rather than routinely sending it to external cloud servers. That is the practical data-control argument for the Dhir approach. Ascendants identifies government departments, educational institutions and organizations handling confidential documents as potential beneficiaries of processing that does not depend on continuous cloud access.

One mission, different deployment goals

The IndiaAI Mission supported development through subsidized access to computing infrastructure and financial assistance for research. That support addresses the substantial computing, infrastructure and research investment required to develop large AI models—requirements that can make it difficult for smaller companies to compete with established international developers.

The initiative’s stated aim is to strengthen domestic AI development around India’s linguistic and industrial requirements.