India pushes its first indigenous AI foundation models into the spotlight
The IndiaAI Mission presented three locally developed foundation models in New Delhi: EKA-1, Dhīr 2B, and Dhīr 5B. The stated goal is clear: to strengthen technological autonomy, including offline and on-site.
The IndiaAI Mission unveiled three indigenous foundation models on November 3 in New Delhi: EKA-1, developed by Soket AI Labs, and Dhīr 2B as well as Dhīr 5B, created by ZenteiQ AiTech Innovations. Designed for uses ranging from public services to cybersecurity, they mark a milestone in India's strategy for AI autonomy. This topic also matters for businesses and administrations, as these models are designed to operate locally without constant reliance on the cloud.
Models calibrated for India, not just for show in a demo
EKA-1 is presented as a sparse hybrid model with 26.3 billion parameters, trained on about 7.5 trillion tokens. It supports the 22 official Indian languages as well as English, with capabilities in mathematics, coding, reasoning, and agentic applications. Soket AI Labs says it was designed for the banking, financial services, and cybersecurity sectors. Abhishek Upperwal, CEO of Soket AI Labs, stated: "With the support of the IndiaAI Mission, we have developed an advanced AI model specifically tailored to India's unique banking, financial services and cybersecurity landscape. It is designed to address sector-specific requirements, understand India's evolving digital ecosystem, and deliver solutions aligned with the country's needs and priorities,"
On the other hand, Dhīr 2B and Dhīr 5B are open-source multilingual assistants intended for on-site and offline deployment in the public sector, defense, industry, and education. ZenteiQ says they support English and 10 Indian languages. The models target document automation, information retrieval, and services for citizens as well as businesses. Prof Sashikumar, CEO of ZenteiQ AiTech Innovations, explained: "Models like Dhīr 2B and Dhīr 5B offer the flexibility to run across different computing platforms, including GPUs and TPUs. Their lightweight architecture also enables on-device deployment on mobile phones, without dependence on cloud connectivity," then: "This allows users to access AI capabilities locally while keeping data on their devices, enhancing privacy, security and accessibility,"
What offline mode changes for public and industrial uses
The choice of local deployment is not trivial. In administrations, armies, factories, or educational institutions, permanent cloud connection is not always possible or desired. Models capable of running on various infrastructures, including mobile, reduce network dependency and keep data closer to its source. It’s less glamorous than a trade show demo, but often much more useful.
All three models were developed by teams selected under the Foundation Models pillar of the IndiaAI Innovation Centre. The public mission provided access to a subsidized national computing infrastructure, as well as support for research expenses. S Krishnan, secretary of the Indian Ministry of Electronics and Information Technology, considered that these indigenous models would strengthen India’s technological autonomy in AI and open productivity gains for government, industry, research, and education. He also mentioned AIKosh, the IndiaAI platform where these models will be made available to researchers, students, and developers.
A first showcase for a local sector still under construction
Sudeep Shrivastava, COO of IndiaAI, presented these releases as a sign of an upgrade in Indian AI, from lightweight systems intended for everyday uses to models designed for more complex enterprise applications. The message is clear: India wants to show it no longer just integrates models from elsewhere but can also produce them for its own linguistic, institutional, and industrial needs. Now remains to see which actors will adopt these systems at scale, and at what pace [to be verified].
By the numbers
- 26.3 billion parameters — EKA-1, according to Soket AI Labs.
- 7.5 trillion tokens — training volume of EKA-1.
- 22 official languages — covered by EKA-1, plus English.
- 10 Indian languages — supported by Dhīr 2B and Dhīr 5B.
- 3 models — presented in New Delhi on November 3 by IndiaAI Mission.