EKA-1 and Dhir AI models released under IndiaAI Mission by Soket, ZenteiQ

The IndiaAI Mission, run by the Ministry of Electronics and Information Technology (MeitY), released three AI foundation models built in India on Tuesday, 3 November 2026. EKA-1, developed by Soket Labs, has 26.3 billion parameters, was trained on about 7.5 trillion tokens and supports all 22 scheduled Indian languages plus English. Dhir 2B and Dhir 5B, developed by ZenteiQ, are smaller open-source models with 2 billion and 5 billion parameters that support English and 10 Indian languages. Both companies were selected under the mission's Innovation Centre pillar for foundation models and received subsidised government computing power.

A foundation model is a large AI system trained on huge amounts of text that can then be adapted for many tasks, from answering questions to writing code. Most of the widely used ones are built by American and Chinese companies and trained mainly on English and a few other major languages. India's IndiaAI Mission is paying for home-grown models partly so that Indian languages are handled well, and partly so that government and businesses are not entirely dependent on foreign systems for sensitive uses.

The two releases show two different strategies. EKA-1 is the larger of the two, built to cover every scheduled language and to handle harder tasks such as reasoning and coding. Soket's chief executive said the model was shaped with India's banking, financial services and cybersecurity needs in mind, sectors where data often cannot leave the country. Its claimed ability to run some tasks on a local device matters for organisations that cannot send data to the cloud.

The Dhir models go the other way: small and open-source, meaning anyone can inspect and adapt them. Smaller models are cheaper to run and can work on modest hardware, which suits chatbots and assistants for schools, local governments and small businesses, and areas where internet connections are unreliable.

These are company claims and early releases. Independent tests of how well the models perform against established systems will matter more than launch descriptions, and details on licences and whether full model weights can be downloaded were not clear in early reports.

Peepals calculation: EKA-1 is about 5.3 times the size of Dhir 5B (26.3 divided by 5) and about 13 times Dhir 2B (26.3 divided by 2). It was trained on roughly 285 tokens for each parameter (7.5 trillion divided by 26.3 billion). EKA-1 covers 23 languages in all (22 scheduled languages plus English), against 11 for Dhir (English plus 10). Our calculation from the reported figures.

Short-term: Researchers, students and start-ups get new Indian-built models to test and build on, including two open-source options.

Long-term: If the models perform well, they could be used in government services, banking and education where Indian-language support and data control matter. Their uptake will be a test of whether the mission's compute subsidies produce models people actually use.

Who is affected: Indian AI developers and start-ups, banks and cybersecurity firms, government departments, and users who need AI tools in Indian languages.

What to watch next: Independent benchmark results for EKA-1 and Dhir, and licence details on the platforms where they are published.

Questions and Answers