India's EKA-1 and Dhīr AI Models Explained

India now has three more homegrown foundation models, and they are not three versions of the same idea. On 3 November 2026, the IndiaAI Mission released EKA-1 from Soket Labs Technology and Research alongside Dhīr 2B and Dhīr 5B from ZenteiQ AiTech Innovations, at an event in New Delhi attended by MeitY Secretary S Krishnan. One is a large model aimed at banks, cybersecurity and complex reasoning in Indian languages. The other two are small models built to run offline, on a laptop, with no internet connection at all. That contrast — big and enterprise-focused versus small and portable — is what makes this launch worth understanding, rather than just another announcement to file away.

Quick answer: On 3 Nov 2026 India's IndiaAI Mission released three homegrown models: EKA-1, a 26.3B-parameter model for 22 Indian languages, and the offline Dhīr 2B and 5B pair.

What was announced on 3 November?

The IndiaAI Mission, the government’s flagship AI programme under the Ministry of Electronics and Information Technology (MeitY), released three foundation models developed by two Indian startups. EKA-1 comes from Bengaluru-based Soket Labs Technology and Research; Dhīr 2B and Dhīr 5B come from ZenteiQ AiTech Innovations, a startup born at the Indian Institute of Science (IISc) Bangalore. Both companies were selected under the IndiaAI Innovation Centre’s Foundation Models pillar, which supports domestic teams building foundation models with subsidised national computing infrastructure and research-expense support.

This is not the mission’s first set of models. The launch adds to earlier IndiaAI foundation models from Sarvam AI, BharatGen and Gnani AI. But it is the first time the mission has released a large enterprise model and a pair of offline-capable small models on the same day — and that pairing is deliberate. As The Hindu BusinessLine reported on 4 November 2026, the models are aimed at very different jobs: EKA-1 at Indian-language enterprise work, the Dhīr pair at secure, document-grounded work in places where connectivity cannot be assumed.

“These three indigenous models are a key milestone under the IndiaAI Mission, showcasing India’s progress in creating homegrown AI technology.” — The Hindu BusinessLine, 4 November 2026

What is EKA-1, and what is it for?

EKA-1 is the heavyweight of the trio: a 26.3-billion-parameter model trained on about 7.5 trillion tokens, covering all 22 scheduled Indian languages plus English. Soket Labs describes it as an open, sparse hybrid language model designed for mathematics, code, reasoning and agentic tasks. Its stated target sectors are banking, financial services and cybersecurity — fields where data sensitivity and local context matter as much as raw capability.

Soket CEO Abhishek Upperwal said the model was tailored to India’s banking, financial services and cybersecurity landscape, aiming to understand sector-specific requirements rather than simply translating a general-purpose model into Indian languages. That framing matters: EKA-1’s pitch is not that it beats the biggest international models on a leaderboard, but that it understands Indian-language enterprise work that those models were never tuned for.

Two honest caveats belong here. First, the launch announcement does not establish that EKA-1 has already been adopted by banks or deployed in operational cyber-defence systems — those are target sectors, not confirmed customers. Second, no public benchmark table was published with the launch, so there is no independent evidence yet of how EKA-1 performs against established alternatives in any of those 22 languages. Parameter counts and training-token totals are useful context, not proof of quality: a 26.3-billion-parameter model can still lag a smaller, better-tuned model on a specific task.

Soket is also building a larger project under the same mission: MeitY selected the company in May 2025 to develop India’s first open-source 120-billion-parameter foundation model tuned for the country’s linguistic diversity, with defence, healthcare and education as target sectors. EKA-1 is the smaller, earlier step on that road.

What are Dhīr 2B and Dhīr 5B for?

The Dhīr pair takes the opposite approach: small, open-source models built to run fully offline on laptops and edge devices. Dhīr 2B is a dense model with 2.03 billion parameters. Dhīr 5B uses a mixture-of-experts design with 5.09 billion total parameters, of which 1.79 billion are active at a time. Both support English and 10 Indian languages, and both are designed for document-grounded automation, knowledge retrieval and citizen or enterprise services in defence, government, industrial and educational settings — particularly in low-connectivity or sensitive environments.

ZenteiQ, led by founder Prof Sashikumaar Ganesan, builds the Dhīr family as scientific foundation models pretrained from scratch rather than adapted from existing model weights. That matters for the use case: an organisation that needs an assistant to read its own documents, on its own machines, without sending anything to a distant cloud, can deploy these models on-premises. The company says the models can be fine-tuned on enterprise data.

A larger Dhīr 9B model is reported to be in development — label that as reported, not launched, because no release date or specification has been published for it.

The offline angle is the Dhīr pair’s real story. Most AI assistants today are thin clients for a faraway data centre; take the connection away and they stop working. A model that runs on a laptop changes who can use AI at all: a district office with patchy connectivity, a defence installation that cannot send data out, a classroom with no reliable internet. Whether the models are accurate enough in each of their 10 languages is a separate question the launch did not answer — offline capability says nothing about accuracy in a particular language, and that will only be settled by independent testing.

How the three models compare

	EKA-1 (Soket Labs)	Dhīr 2B (ZenteiQ)	Dhīr 5B (ZenteiQ)
Size	26.3 billion parameters	2.03 billion parameters (dense)	5.09 billion total / 1.79 billion active (MoE)
Training data	About 7.5 trillion tokens	Not stated	Not stated
Languages	All 22 scheduled Indian languages + English	English + 10 Indian languages	English + 10 Indian languages
Licence / access	Open, sparse hybrid model	Open-source	Open-source
Runs offline?	Not the pitch — enterprise deployment	Yes, on laptops and edge devices	Yes, on laptops and edge devices
Built for	Banking, cybersecurity, maths, code, reasoning, agentic use	Document-grounded work in low-connectivity or sensitive settings	Document-grounded work in low-connectivity or sensitive settings

Who can actually use them?

The models are intended for researchers, students and the public, with access expected through platforms such as AIKosh, the IndiaAI Mission’s data and model platform. Do not treat that as a promise of a one-click download: at the time of writing, no verified public download page or model-card link had been confirmed for these three models, and this article will not link to one until it exists. The honest position is that the launch is a release under a government mission, not a consumer chatbot launch — there is no user-facing app to try today, and no pricing, because these are foundation models for builders, not products for end users.

That distinction is also what separates this story from yesterday’s Sarvam Prism piece on this site. Sarvam is a private company’s decision-making model with a published benchmark story; EKA-1 and Dhīr are government-mission models whose value will be proven — or not — by what Indian organisations build on them in the coming months.

Our take: five questions that decide whether this launch matters

Sovereign-AI launches are easy to applaud and hard to judge. Here is the five-question test we will use to follow this one — and that readers can use for any national-model announcement:

Are the weights actually available? A model you cannot download is a press release, not a tool. Watch for a public model card and weights, not just an announcement.
Are there independent benchmarks? Company-reported numbers are a starting point. Independent, reproducible results in Indian languages are what prove the language claim.
Is anyone named using it? Target sectors are not customers. Named deployments — a bank, a department, a university — are the real signal.
Does the offline claim survive contact with reality? Dhīr’s promise is specific and testable: run it on a laptop, in a low-connectivity setting, on real documents. That test can be run by anyone once the weights are out.
What gets built on top? Foundation models are infrastructure. The launch matters if Indian developers ship useful things on EKA-1 and Dhīr that they could not ship on imported models.

Apply the test today and the honest score is one out of five: the models exist and have been released under a named government programme, which is real. The other four answers are still open. That is not a criticism — it is the normal state of a launch two days old. The follow-up is what counts, and we will track it.

What to watch next

Three things, in order. First, the weights and model cards: whether EKA-1 and the Dhīr pair appear on a public hub with licences and documentation. Second, independent benchmarks in Indian languages — the claim that matters most for EKA-1 has no independent evidence yet. Third, named deployments: the first bank, government department or university that puts one of these models into real use will tell us more than any parameter count. The larger Dhīr 9B, reported to be in development, is a fourth marker worth watching, but it is not a launch until it is announced.

Frequently Asked Questions

What is EKA-1?

EKA-1 is a 26.3-billion-parameter foundation model developed by Soket Labs under the IndiaAI Mission, trained on about 7.5 trillion tokens. It covers all 22 scheduled Indian languages plus English and is aimed at banking, cybersecurity, mathematics, code, reasoning and agentic tasks.

What are Dhīr 2B and Dhīr 5B?

They are two open-source foundation models from ZenteiQ AiTech Innovations, also released under the IndiaAI Mission on 3 November 2026. Dhīr 2B has 2.03 billion parameters and Dhīr 5B has 5.09 billion total parameters (1.79 billion active). Both support English and 10 Indian languages and are built to run offline on laptops and edge devices.

Can I download EKA-1 or Dhīr today?

Not through a verified public link at the time of writing. The models are intended for researchers, students and the public, with access expected through platforms such as AIKosh, but no confirmed download page had been published when this article was written. Do not trust unofficial download links.

Are EKA-1 and Dhīr the same as Sarvam’s models?

No. Sarvam AI is a private company whose models, including Sarvam Prism, were released under the same IndiaAI Mission but are a separate story. EKA-1 comes from Soket Labs and Dhīr from ZenteiQ; all three are distinct models from different builders.

Do these models have benchmark results?

No public benchmark table was published with the launch. All performance claims so far are company- or government-reported, and independent results in Indian languages are still awaited.

Sources and methodology

Sources: The Hindu BusinessLine, “IndiaAI Mission releases indigenous foundation models EKA-1, Dhīr 2B and Dhīr 5B” (4 November 2026); The Hindu BusinessLine, “IndiaAI Mission: Indian startups launch offline AI models Dhīr and EKA-1” (4 November 2026); CIO / Elets, “IndiaAI Foundation Models: EKA-1 and Dhir Launch” (3 November 2026); AI Affairs, “IndiaAI Mission releases EKA-1 and two Dhīr foundation models” (4 November 2026). Internal links: our Sarvam Prism explainer, our Muse AI availability in India guide, and our Kimi K3.1 Pro open-weights guide.

Method: desk research on 5 November 2026, cross-checking the launch facts across four independent reports of the same event; every figure is attributed to the company or government that reported it, and anything not yet published (benchmarks, deployments, download links) is stated as unproven rather than assumed. This article was not hands-on tested — there is no public model to test yet.