India has launched three home-grown AI models under the IndiaAI Mission, including two small models designed to run on phones and laptops without an internet connection. The AI models were unveiled in New Delhi on 3 November 2026 by the Ministry of Electronics and Information Technology, in the presence of ministry secretary S. Krishnan and IndiaAI Mission chief operating officer Sudeep Shrivastava. The launch pairs two different designs: a large model built for complex work in banking and cybersecurity, and two compact models built for everyday use in Indian languages. All three were developed by Bengaluru startups under the mission’s Foundation Model programme.
What the three AI models do
Soket Labs developed the largest of the three AI models, EKA-1, an open model with 26.3 billion parameters trained on about 7.5 trillion tokens. It is built for mathematics, coding and reasoning, and for work in banking, financial services, cybersecurity and enterprise automation. EKA-1 can also carry out some of its work directly on a local device, such as a smartphone, without mobile data or Wi-Fi. That helps in places where connectivity is patchy, and in settings where sending data to a remote cloud is not wanted. Soket Labs plans to make the model weights openly available within a month, with early access through a waitlist.
ZenteiQ AiTech Innovations developed the other two, Dhīr 2B and Dhīr 5B. Both are open-source, lightweight multilingual assistants that support English and 10 Indian languages. Dhīr 2B is a dense model with 2.03 billion parameters. Dhīr 5B uses a mixture-of-experts design with 5.09 billion parameters in total, of which 1.79 billion are active at a time. The two AI models are compact enough to run fully offline on laptops and edge devices, and on phones with at least 6 GB of memory. They can also be deployed on an organisation’s own infrastructure as secure on-premise assistants, for document work, knowledge retrieval and citizen services in defence, government, industry and education.
AI models built for Indian languages
The emphasis on Indian languages runs through the launch. Most large AI systems are trained mainly on English and a handful of other languages, which leaves users working in Indian languages with weaker results. The new AI models were built to work in English and 10 Indian languages from the start, and to run on ordinary devices rather than in a distant data centre. ZenteiQ was born at the Indian Institute of Science in Bengaluru and is led by founder Prof. Sashikumaar Ganesan. Ganesan said the models need no network connection and can run on a phone with at least 6 GB of memory. The model weights are already available through AIKosh, the government’s platform for AI data and models.
The three AI models form part of a wider government push to build AI in India rather than buy it from abroad. The IndiaAI Mission was approved by the Union Cabinet in March 2024 and runs through seven pillars, including computing capacity, data platforms and support for home-grown foundation models. The ministry framed the launch as a step towards domestic capability in defence, banking and government work, where data often cannot leave the country. Reducing reliance on foreign technology was the stated aim, and the two startups were picked under the mission’s Innovation Centre to build models for Indian conditions.
How the AI models will be judged
What the launch does not yet show is how the three AI models perform against established international systems. The government has not supplied comprehensive independent benchmark results, and detailed deployment figures are not yet available. Parameter counts and training-token totals give a sense of scale, not proof of quality. Offline capability says nothing about accuracy in a particular language, and a model that runs on a phone still has to answer well in the language the user speaks. The announcement also gave no pricing, and no bank or defence user has been named as an early adopter. For airgapped systems, where a network is cut off from the outside entirely, the design is the point: the work stays on the device, and the data does not travel.
Soket Labs plans a 9-billion-parameter model in December for researchers and engineers working on specialised problems. The Dhīr weights are already on AIKosh, and EKA-1 is due to follow through a waitlist within a month.