IndiaAI Mission Launches EKA-1, Dhīr AI Models Built for India

MeitY's IndiaAI Mission has released EKA-1 and two Dhīr models — open, multilingual AI built for Indian banking, defence and offline use.

Nov 18, 2026 - 07:10
4 min read
 0
IndiaAI Mission Launches EKA-1, Dhīr AI Models Built for India

Lose your internet connection in a village in Bihar or a bank branch in a tier-3 town, and most AI chatbots stop dead. That's the exact gap a trio of newly released Indian-made AI models is aiming to close — and they come straight out of the government's own AI mission, not a Bengaluru startup's pitch deck.

What actually got released

In early November, the IndiaAI Mission — the Ministry of Electronics and IT's (MeitY) flagship AI program — put out three new "foundation models." A foundation model is just a general-purpose AI model that other apps and tools get built on top of, the same role GPT or Gemini play for products you already use.

The headline release is EKA-1, built by Gurugram-based Soket Labs. It's a 26.3-billion-parameter model (parameters are roughly the number of internal knobs a model tunes while learning — more isn't automatically better, but it's a rough proxy for capability) trained on about 7.5 trillion tokens of text. It covers all 22 languages listed in the Indian Constitution's Eighth Schedule, plus English, and is pitched specifically at banking, financial services and cybersecurity use cases, alongside coding and reasoning tasks.

Alongside it came Dhīr 2B and Dhīr 5B from ZenteiQ — smaller, open-source multilingual models built to run on-premise or fully offline, including on phones, for government, defence, industrial and education deployments. All three are being made available through AIKosh, the government's open AI data and model repository, rather than locked behind a single company's paid API.

Why "sovereign AI" is the real story here

None of this happened in isolation. New Delhi has been pushing the idea of "sovereign AI" — models trained, hosted and controlled within India rather than rented from a handful of US or Chinese firms — for over a year now, backing proposals from outfits like Sarvam AI and BharatGen alongside Soket and ZenteiQ. The pitch isn't nationalism for its own sake; it's about not being stuck mid-project if a foreign provider changes its pricing, its terms of service, or its willingness to serve Indian customers at all.

There's also a cost angle that matters more than it sounds. Every API call to a foreign large language model is, functionally, a small recurring export of money and data. Models that can run locally — even smaller, less flashy ones like the Dhīr pair — cut both the bill and the dependency.

A model that only works with a live internet connection isn't useful to the India that still runs on patchy 4G and shared bank-branch terminals — it's useful to the India that already has everything else.

What it means for Indian developers and companies

For Indian developers and startups, open weights matter more than the press-release numbers. "Open weights" means the actual trained model file is downloadable and runnable on your own servers, not just accessible through someone else's paid API — so a fintech or ed-tech startup can fine-tune EKA-1 or Dhīr for its own use case without sending customer data to a server outside the country. That's not a minor detail under the Digital Personal Data Protection (DPDP) Act, which puts real constraints on how personal data of Indian users can be processed and where. A startup building, say, a vernacular customer-support bot for a regional bank now has a local option that doesn't require a cross-border data transfer conversation with its compliance team.

It's also a genuine opening for smaller players who can't afford frontier-model API bills at scale. A few things worth tracking as this rolls out:

  • Whether EKA-1's full weights actually ship on the promised timeline, or whether access stays limited to a waitlist
  • How the 22-language claim holds up in practice — multilingual benchmarks for low-resource Indian languages have historically lagged the headline number
  • Whether banks and government bodies actually deploy these models in production, or whether this stays a pilot-stage showcase

The harder part comes next

Releasing a model is the easy half. The IndiaAI Mission has also been racing to add GPU capacity — tens of thousands of chips empanelled over the past year — because training and running models like EKA-1 at scale needs serious compute, and India still largely rents that compute rather than owning it outright. The real test for EKA-1 and the Dhīr models isn't the launch announcement; it's whether banks, state governments and startups actually build on them a year from now, instead of quietly going back to whichever foreign API worked fine all along.

Short URL: https://code24.in/021cbc9c

What's Your Reaction?

Like Like 0
Dislike Dislike 0
Love Love 0
Funny Funny 0
Angry Angry 0
Sad Sad 0
Wow Wow 0
Ashif Sadique As an full-stack developer, I'm passionate about sharing tutorials and tips that aid other programmers. With expertise in PHP, Python, Laravel, Angular, Vue, Node, Javascript, JQuery, MySql, Codeigniter, and Bootstrap. To me, consistency and hard work are the keys to success.