India's EKA-1 and Dhīr Want AI That Works Offline, in 22 Languages — Explained Honestly

Most AI models assume you are online, speaking English, and running on someone else's cloud. India's new foundation models start from the opposite assumption: patchy connectivity, 22 scheduled languages, and work that cannot always leave a private server — or a phone. On 4 November 2026, the IndiaAI Mission, under India's Ministry of Electronics and IT (MeitY), released three indigenous foundation models: EKA-1, a large model from Soket Labs Technology and Research (also written Socket AI), and Dhīr 2B and Dhīr 5B, two compact models from ZenteiQ AiTech Innovations.

That is a genuinely interesting direction, and it is also a launch announcement — not a finished product you can download and run today in every case. This article keeps those two things separate throughout: what the announcement actually says, what you can access today, and what is only planned or claimed so far. Every figure below comes from the announcement and the reporting that followed, checked 4 November 2026. We have not tested any of these models hands-on, and nothing here is a hands-on finding.

What it is

Three models, two very different jobs:

• EKA-1 (Soket Labs Technology and Research): 26.3 billion parameters, trained on about 7.5 trillion tokens, described as an open sparse hybrid model. It covers all 22 scheduled Indian languages plus English, and is designed for math, coding, reasoning and agentic use — and, for Indian languages specifically, banking and cybersecurity work. It is built to run on private servers or a private cloud, and to do some on-device inference without mobile data or WiFi.
• Dhīr 2B and Dhīr 5B (ZenteiQ AiTech Innovations): compact, open-source models covering English plus 10 Indian languages, designed to run directly on phones without internet. Their stated purpose is offline-capable, document-grounded automation and knowledge retrieval — citizen and enterprise services in defence, government, industrial and educational settings, especially where connectivity is low or the environment is sensitive.

The shared idea is worth naming plainly: AI that does not depend on a constant connection or on sending your data to a distant data centre. EKA-1 aims at the private-server end of that idea; Dhīr aims at the phone-in-your-pocket end. If either works as described, it matters far beyond India — for anyone building for low-connectivity users, regulated data, or languages that the big English-first models still serve poorly.

What you can actually access today — and what is only planned

• Dhīr 2B and Dhīr 5B: weights already available, per the reporting, through AIKosh. These are the two models described as open-source. If you want to try something from this announcement today, Dhīr is the place to start — with the caveat that we have not run them ourselves.
• EKA-1: NOT openly available today. Its weights are to be made openly available within a month of the announcement, and initial access is via a waitlist. Do not read "open" in descriptions of EKA-1's architecture as meaning "downloadable now" — as of 4 November 2026, it is waitlisted, not released in the downloadable sense. We have not stated a licence for EKA-1 here, because the sources used for this article do not give one.
• A Dhīr 9B model is planned for December 2026, aimed at researchers and engineers. That is a plan, not a release.
• No benchmarks, prices, or API availability are reported here, because the sources used for this article do not give them. Any article that quotes an EKA-1 or Dhīr benchmark score, a price, or an API endpoint today is adding something the announcement reporting we checked does not contain.

Why it matters — if the access and the quality follow through

The language coverage is the headline for a reason. A model that genuinely works across all 22 scheduled Indian languages plus English (EKA-1) — or across English plus 10 Indian languages on a phone with no connection at all (Dhīr) — addresses users whom the current generation of AI tools serves worst: people whose first language is not English, and people whose connection is not reliable. Banking and cybersecurity are named as EKA-1 use areas for Indian languages, and citizen services as a Dhīr use area; in both cases the point is the same — sensitive or everyday work done closer to the user, not shipped off to a cloud.

The offline angle is the second reason to pay attention. "Runs on a phone without internet" is a claim with a long history of asterisks — it depends on the phone, the task, and how much capability survives without the cloud — and Dhīr's compact 2B and 5B sizes are consistent with that ambition without proving it. Likewise, "some on-device inference without mobile data or WiFi" for EKA-1 is a deliberately partial claim: some inference, not the whole model running fully offline on a handset. Read both claims at exactly the strength stated, no stronger.

The publisher's claims — labelled as the publisher's

Every specification in this article — 26.3 billion parameters, about 7.5 trillion training tokens, 22 scheduled languages plus English, the math/coding/reasoning/agentic design goals, the banking and cybersecurity focus, the private-server and private-cloud deployment, the on-device inference without mobile data or WiFi, the Dhīr sizes and language counts, the offline phone operation, the document-grounded automation and knowledge retrieval, the defence/government/industrial/educational sectors, and the December 9B plan — comes from the IndiaAI Mission announcement as reported by The Hindu, Hindu BusinessLine, Economic Times Telecom and ANI (via Punjab Kesari) on 4 November 2026. We have not run these models, reproduced any of these figures, or tested them hands-on. Treat them as the publishers' and builders' own claims, not as independently verified results.

On pricing and licensing, there is nothing to report beyond what is stated above: Dhīr is described as open-source, and no price, tier or commercial offering is given in the sources used here. Nothing here should be read as implying a cost for using any of these models, and no licence terms beyond the open-source description of Dhīr have been invented.

Who it matters for

• Developers building for India or for low-connectivity users. Dhīr 2B and 5B are the immediately testable part of this announcement: small, open-source, phone-sized models for offline, document-grounded work. If you serve users with patchy connectivity, try them on your own documents and tasks before believing any capability claim — including the ones in this article.
• Teams in banking, government, defence, industry and education. These are the sectors named in the announcement. The relevant question is not whether the models exist — they have been announced — but whether they handle your documents, your languages and your security requirements. EKA-1's private-server story is aimed squarely at that question, but with weights still waitlisted, it cannot be answered hands-on yet.
• People working in Indian languages. EKA-1's 22-language coverage and Dhīr's 10-language coverage are the most consequential claims here if they hold up in practice. Language coverage on a spec sheet and quality in a real conversation are different things; the gap between them is what testing is for.
• Anyone watching sovereign and offline AI. India is not the only country pursuing home-grown models for its own languages and infrastructure, and the offline-phone direction is one the whole industry is moving toward. This announcement is a useful data point in that trend — three models, one mission, one day — not a verdict on it.

The limits, stated plainly

We have not tested EKA-1, Dhīr 2B or Dhīr 5B hands-on, and nothing here is a hands-on finding. Everything above comes from the 4 November 2026 announcement reporting, checked the same day.

• EKA-1 is not downloadable today. Weights are promised within a month, with initial access via waitlist. Until they appear, every EKA-1 capability claim is unverified by anyone outside the builders.
• Dhīr's open-source availability is reported via AIKosh, but we have not verified the download, the licence text, or the models' behaviour. "Available" is the reporting's word, not our test result.
• No benchmarks are reported here, because none were given in the sources used. There is no score in this article to compare against any other model, and that absence is deliberate.
• No prices, API endpoints or licence details beyond Dhīr's open-source description are reported, because the sources used do not give them. Do not assume a price of zero, a public API, or any particular licence for EKA-1.
• "Offline" is partial and conditional. Dhīr is described as running directly on phones without internet; EKA-1 is described as doing some on-device inference without mobile data or WiFi and running on private servers or a private cloud. Neither description promises that every capability works fully offline on every device.
• The language counts are coverage claims, not quality claims. Covering 22 languages (EKA-1) or 10 Indian languages plus English (Dhīr) says nothing by itself about fluency, accuracy or safety in any one of them.
• A 9B Dhīr model for December 2026 is planned, not released. Do not plan around it until it appears.
• No hands-on testing was performed for this article, and no benchmarks, prices, API details or licence terms have been invented to fill gaps.

The bottom line is a measured one. Three indigenous models in one day, aimed at the hardest parts of real-world AI deployment — many languages, poor connectivity, private infrastructure — is a serious announcement, and Dhīr's reported availability through AIKosh gives developers something to actually try. But the biggest model in the set is still behind a waitlist, no benchmark or price exists in the reporting to judge any of them by, and "offline" and "multilingual" are claims that only hands-on testing can confirm. Watch for EKA-1's weights within the promised month, and for Dhīr's real-world behaviour on phones and documents: those two events, not this announcement, will tell us whether India's offline, multilingual AI bet pays off. Until then, treat it as exactly what it is — a promising launch, honestly reported, with its limits in the open.

Sources: The Hindu, 4 November 2026 (https://www.thehindu.com/sci-tech/technology/indiaai-mission-releases-indigenous-foundation-models-eka-1-dhir-2b-and-dhir-5b/article71663984.ece); Hindu BusinessLine, 4 November 2026 (https://www.thehindubusinessline.com/info-tech/indiaai-mission-releases-indigenous-foundation-models-eka-1-dhir-2b-and-dhir-5b/article71663692.ece); Economic Times Telecom, 4 November 2026 (https://telecom.economictimes.indiatimes.com/news/internet/indiaai-mission-releases-indigenous-foundation-models-eka-1-dhr-2b-dhr-5b/135520120); ANI via Punjab Kesari, 4 November 2026 (https://english.punjabkesari.com/business/indiaai-missions-three-indigenous-ai-models-target-offline-mobile-access-secure-use-in-critical-sectors/). All specifications are the publishers'/builders' own claims, not independently tested results. No hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links.

Frequently asked questions

What did the IndiaAI Mission release on 4 November 2026?
Three indigenous foundation models: EKA-1, by Soket Labs Technology and Research (also written Socket AI), and Dhīr 2B and Dhīr 5B, by ZenteiQ AiTech Innovations. The release was made under the IndiaAI Mission, under India's Ministry of Electronics and IT (MeitY).

What is EKA-1?
A 26.3-billion-parameter model trained on about 7.5 trillion tokens, described as an open sparse hybrid model, covering all 22 scheduled Indian languages plus English. It is designed for math, coding, reasoning and agentic use, and for Indian-language banking and cybersecurity work, running on private servers or a private cloud, with some on-device inference without mobile data or WiFi. These are the builders' claims as reported, not independently tested results.

Can I download EKA-1 today?
No — not as of 4 November 2026. Its weights are to be made openly available within a month of the announcement, and initial access is via a waitlist. Do not call EKA-1 open-source today: it is waitlisted, not openly released.

What are Dhīr 2B and Dhīr 5B?
Compact, open-source models from ZenteiQ AiTech Innovations, covering English plus 10 Indian languages, designed to run directly on phones without internet, for offline-capable, document-grounded automation and knowledge retrieval in citizen and enterprise services — including defence, government, industrial and educational settings, especially low-connectivity or sensitive environments. Their weights are reported as already available through AIKosh.

What is the difference between EKA-1 and Dhīr?
Scale and deployment: EKA-1 is the large model (26.3 billion parameters) aimed at private servers and private cloud, with some on-device inference; Dhīr 2B and 5B are small, phone-sized models aimed at running directly on a handset with no internet. EKA-1 covers 22 scheduled Indian languages plus English; Dhīr covers English plus 10 Indian languages. EKA-1 is waitlisted; Dhīr is reported as available now.

Is there a bigger Dhīr model coming?
A 9B Dhīr model is planned for December 2026, aimed at researchers and engineers. It is a plan, not a release.

Are there benchmarks or prices for these models?
None are reported in this article, because the sources used for it do not give any. No benchmark score, price, API endpoint or — beyond Dhīr's open-source description — licence detail has been invented to fill that gap.

Were these models tested hands-on for this article?
No. No hands-on testing was performed; all facts and figures come from the 4 November 2026 announcement reporting listed in the sources, and every specification is the publishers'/builders' own claim, not an independently tested result.