India has released three indigenous artificial intelligence foundation models under the IndiaAI Mission, in a step the government presents as a milestone in building homegrown AI capability for Indian languages and for secure, offline use. The models, EKA-1 and the DhIr 2B and DhIr 5B pair, were unveiled in New Delhi on 3 November by Electronics and Information Technology Secretary S. Krishnan, and the announcement was reported widely on 4 November.
EKA-1, developed by Soket Labs Technology and Research, is the larger of the three. It is a foundation model with 26.3 billion parameters, trained on approximately 7.5 trillion tokens, and it supports all 22 scheduled Indian languages alongside English. Its intended uses span mathematics, coding, reasoning and agentic applications, as well as sectors such as banking and financial services, cybersecurity and enterprise automation. For a country where most AI systems are still trained primarily on English, a model built from the start for India’s linguistic diversity is a significant proposition.
The DhIr models, developed by ZenteiQ AiTech Innovations, are smaller and deliberately different in purpose. DhIr 2B and DhIr 5B are open-source models supporting English and 10 Indian languages, designed as secure, on-premise assistants that can run without a continuous internet connection. Their target environments are precisely those where cloud-based AI is hardest to use: defence, government and education offices, industrial sites and areas with limited connectivity, and any setting where documents are sensitive and cannot leave the premises.
Both companies were selected under the Foundation Models pillar of the IndiaAI Innovation Centre, the programme through which the government supports domestic teams building large language models. The government has said the teams received access to subsidised national AI computing infrastructure and support for related research expenses, which made the development of the models possible. The IndiaAI Mission itself was approved by the Union Cabinet in March 2024 and operates through seven pillars covering compute, foundation models, datasets, applications, skills, startup financing and safe and trusted AI.
The practical question is what these models will actually be used for. The government says they are intended to be accessible to researchers, students, innovators and the broader public through platforms such as AIKosh, the IndiaAI datasets platform. EKA-1’s multilingual reach makes it a candidate for citizen services, customer support and content in Indian languages, while the DhIr models’ offline capability makes them suited to document-grounded retrieval: answering questions from an organisation’s own files without sending those files to an external server.
There are limits to what the launch, by itself, proves. Foundation models are judged over time on benchmarks, accuracy and developer adoption, not on announcement day. The DhIr models are built from scratch rather than adapted from existing weights, giving India full control over the technology. EKA-1, at 26.3 billion parameters, is mid-range among current models but large enough for serious enterprise and research use.
The launch fits a broader policy arc: reducing dependence on foreign AI infrastructure while keeping the ecosystem open, through the mission’s subsidised compute programme. Homegrown models can be tuned to Indian languages and administrative contexts in ways imported models cannot easily match.
For students, researchers and startups, the immediate benefit is access: three new models, two of them open-source, built for the languages and connectivity conditions of India. Whether they become the backbone of Indian-language AI or remain one more announcement in a crowded field will depend on what developers build with them in the months ahead. The release on 3 November is, at minimum, the moment when India’s foundation-model programme moved from selection and funding to models that the public can actually use.
