India's IndiaAI Mission has three new home-grown foundation models. On 3 November, the Ministry of Electronics and Information Technology (MeitY) unveiled EKA-1, Dhīr 2B and Dhīr 5B, all built by Bengaluru companies under the mission's Foundation Model initiative.
What was released
EKA-1, from Socket AI, is a 26-billion-parameter model aimed at mathematics, coding and reasoning. It was trained on 7.5 trillion tokens and supports more than 100 Indian and international languages, plus programming languages. Socket AI built it to handle multi-step tasks and execute code, and to run on-premises, in a private cloud or in an air-gapped environment.
Dhīr 2B and Dhīr 5B, from ZenteiQ, are mixture-of-experts models trained from scratch on Google Cloud TPUs. They cover English and 10 Indian languages and run directly on tested mobile hardware with no internet connection. ZenteiQ is releasing the weights under the Apache 2.0 licence, with technical documentation and evaluation tools.
All three are available through AIKosh, the government's AI repository, to researchers, students, startups, enterprises and the public.
Why the air-gap detail matters
MeitY Secretary S. Krishnan framed the launch around strategic autonomy, but the more useful remark was about deployment. He said these models can be run in air-gapped environments, and that this could open up enterprise AI, which in his words "thus far hasn't really taken off."
That is a familiar problem in Indian enterprises. Model quality is rarely what stops a project. Where the model runs, and where the data goes, usually is. Banks, insurers, hospitals and government departments often cannot send sensitive data to a hosted API outside their control. A capable model that runs inside the firewall, or on a device with no network at all, takes that objection off the table.
What it doesn't settle
Small and mid-sized models are not frontier models, and the announcement doesn't include benchmark comparisons against them. The real test is whether enterprises can take EKA-1 or Dhīr, fine-tune them on their own data and get something that holds up in production. An Apache 2.0 licence helps, because it lets teams do that without negotiating terms.
If you run technology for an enterprise, the next step can be small. Pick one internal use case where data residency has blocked an AI pilot and try one of these models on it.
Source: MeitY unveils indigenous AI Models EKA-1, Dhīr 2B and 5B under IndiaAI mission, The Tribune (ANI)
