India has taken a significant step towards building its own artificial intelligence capability with the unveiling of three homegrown foundation models in New Delhi. The launch, held on 3 November 2026 under the IndiaAI Mission, marks the country’s most concrete attempt yet to reduce its dependence on foreign AI systems and to build models designed from the start for Indian languages and Indian users.
The headline model is EKA-1, developed by Bengaluru-based Soket AI Labs. It has 26.3 billion parameters and was trained on about 7.5 trillion tokens, placing it among the larger models built in India to date. EKA-1 supports English and all 22 languages listed in the Eighth Schedule of the Constitution, a breadth of language coverage that sets it apart from most global models, which remain strongest in English and a handful of other major languages. It is released under the Apache 2.0 licence and is listed on AIKosh, the government’s AI platform, with restricted access, registration and a waitlist.
Alongside EKA-1, the launch introduced two smaller models from ZenteiQ AiTech Innovations, named Dhir 2B and Dhir 5B. These have about two billion and five billion parameters respectively, and support English and 10 Indian languages. Their defining feature is size: they are designed to run offline on laptops and phones with at least six gigabytes of memory, bringing AI capability to devices and settings where a constant internet connection or a powerful server cannot be assumed.
The launch was attended by S. Krishnan, Secretary of the Ministry of Electronics and Information Technology, along with IndiaAI officials. Their presence underlined the government’s view that foundation models are a matter of national capability, not simply a commercial product. Access to the models is being provided through AIKosh, and Soket has said it plans a nine-billion-parameter model in December, with open weights to follow within a month.
The models were built under the Foundation Models pillar of the IndiaAI Mission, which was approved in March 2024 with an outlay of more than 10,371 crore rupees. The mission’s premise is that India should not have to rely entirely on models built elsewhere, trained overwhelmingly on English-language data from the West. A model that understands India’s languages, and can run on ordinary devices, could in principle serve classrooms, clinics, government offices and small businesses that global models reach only imperfectly.
It is important to be clear about what has not yet been shown. No comprehensive independent benchmark results have been published for these models, so there is no verified basis on which to compare their performance with established global systems. The launch establishes that the models exist and are available for use, but it does not by itself prove how well they perform on standard tests of reasoning, knowledge or language ability. That evidence will only come when independent evaluations are carried out and published.
For now, the significance of the announcement lies in the direction it signals. India has moved from talking about sovereign AI to releasing models built in the country, under a government mission, with Indian languages at their centre. Whether EKA-1 and the Dhir models can compete on quality with the best systems in the world remains an open question, and one that only independent testing can answer. But the launch itself is a milestone, and the next test will be whether developers, researchers and ordinary users take these models up and put them to work.
