Two new Indian foundation models launched late on Tuesday 3 November 2026 under the IndiaAI Mission show that serious AI no longer has to live in the cloud. ZenteiQ AiTech Innovations launched Dhīr, an offline-capable model family, and Soket Labs Technology and Research launched EKA-1, a large multilingual model. For businesses, the headline is the pattern: AI that keeps working when the internet does not, and keeps sensitive data on the device. If your teams work in warehouses, clinics, branches or rural sites, that pattern matters.

This guide separates verified launch facts from the companies' own claims, explains when offline AI beats the cloud, what on-device deployment demands, and how to pilot one offline workflow, as in our AI agent development and AI automation work.

Dhīr and EKA-1: Offline AI Models That Work Without the Internet

What was actually launched

Indian startups ZenteiQ AiTech Innovations and Soket Labs launched indigenous foundation models under the IndiaAI Mission late on Tuesday evening, in the presence of S Krishnan, Secretary of MeitY. They join earlier IndiaAI Mission models from Sarvam AI, BharatGen and Gnani AI, accessible via platforms such as AIKosh.

ZenteiQ's Dhīr comes in two sizes, Dhīr 2B and Dhīr 5B. The company describes them as open-source, offline-capable models for document-grounded automation and knowledge retrieval in defence, government, industrial and educational settings. It says Dhīr runs as a secure on-premises multilingual assistant (English and 10 Indian languages) on GPUs, TPUs and any smartphone with at least 6GB of RAM, without cloud dependence, keeping data on device. Those are company claims, not verified facts — test them on your own devices first.

Soket's EKA-1 is described by the company as a 26.3-billion-parameter open sparse hybrid model trained on around 7.5 trillion tokens, spanning all 22 scheduled Indian languages plus English, designed for maths, code, reasoning and agentic use, and targeted at banking, financial services and cybersecurity. These are company claims reported at launch, not proven benchmark results.

For the full launch report, see the Hindu BusinessLine report on the IndiaAI Mission offline models. An offline model still needs a well-designed assistant around it — see chatbot development options.

Why offline AI matters far beyond India

Connectivity is uneven everywhere, and data rules are tightening everywhere. A driver in a basement car park or a nurse on a ward with patchy Wi-Fi faces the same problem: the cloud model that works at headquarters stutters exactly where the work happens. Offline AI removes that failure point for the tasks it can handle, and keeps data closer to its source.

The same logic applies across Coheser's markets: the United Kingdom and the USA, where regulated work often restricts where data may travel; the UAE and Saudi Arabia, where large facilities and remote sites routinely outrun their connectivity; and Pakistan, where branches and rural sites face the same pressures. This article does not claim Dhīr or EKA-1 are available in Pakistan or tested there — the transferable lesson is the offline pattern, whatever model a business ultimately deploys.

When offline beats cloud — and when it does not

Offline is not better in general — only for specific jobs. Match the job to the deployment, not the trend.

Offline usually wins when:

• Connectivity is intermittent, slow or absent where the work happens.
• The task is bounded and repeatable, such as answering from fixed documents or following a checklist.
• Data is sensitive and should not leave the device or the premises.

Cloud usually still wins when:

• The task needs the largest model, or knowledge that changes by the hour.
• Many users must share one live, centrally updated source of truth.

The middle ground is a hybrid: an on-device model handles routine offline-capable work, and hands over to a cloud model when the question is complex or high-stakes. That mirrors a good retrieval-augmented chatbot, which answers from your documents rather than guessing — see our guide to AI automation for logistics companies.

What on-device deployment really requires

"Works offline" hides real engineering. Four requirements decide whether a pilot succeeds or stalls.

Devices and model size

Model size is a trade-off: a smaller model such as Dhīr 2B runs more easily on ordinary phones, while a larger one may answer better but needs more memory, storage and battery. Test on the actual handsets your team carries, including older ones.

Updates without the cloud

An offline model is frozen when installed. Documents, prices and procedures change, and a model that cannot be updated gives confident wrong answers. Plan the update route before the pilot: devices syncing when back in connectivity, and a way to pull a bad version.

Data and document grounding

Dhīr is pitched as document-grounded: the assistant answers from your manuals, price lists and procedures, not from general training data. That only works if your documents are clean, current and permissioned. Start with one well-maintained document set, and record which document each answer came from.

Security and support

Keeping data on device helps privacy, but devices can be lost or stolen, so encryption, device management and remote wipe still matter. On-premises deployment also means your team or vendor owns uptime and troubleshooting — price that in. Many businesses therefore pair a pilot with custom software development and Flutter app development, so assistant, device controls and sync arrive as one supported product.

Where the offline pattern fits first

The strongest first candidates are jobs where staff already carry a phone or tablet and consult documents: field engineers checking a manual, warehouse staff confirming a procedure, clinic teams answering standard queries, or branch staff retrieving policy details. Our restaurant ERP work shows the same principle: software helps most when it works during the rush. Likewise, an AI calling agent must keep performing when volumes spike. Pilot where connectivity is worst, and prove the model with staff before customer-facing use.

How to pilot one offline workflow

A disciplined pilot takes weeks, not months:

1. Pick one workflow. A single, frequent, document-based job with a clear success measure, such as time saved. Nothing safety-critical first time.
2. Fix the documents. One document set, outdated versions removed, one named owner for updates.
3. Test on real devices. The oldest handsets in circulation, in the actual locations. Record speed, accuracy and battery impact.
4. Run it with staff first. A small team, every wrong answer collected and fixed before widening access.
5. Define the handover. Exactly when the assistant escalates to a human or cloud model — visible, not a dead end.
6. Measure and decide. Compare against the current way of working, then expand, adjust or stop.

The same discipline underpins successful AI agents for small business automation, and a WhatsApp bot can cover the online side of the same workflow.

Risks and honest limits

Offline models are not a shortcut around good data, good design or good governance. A small on-device model will be wrong more often than a large cloud model on open-ended questions, and its mistakes can sound just as confident. Multilingual claims deserve scrutiny: supporting a language is not the same as answering accurately in it, so test each language separately. EKA-1's breadth across 22 scheduled Indian languages plus English is exactly why independent testing matters more than launch-day coverage.

There is also a maintenance trap: an offline deployment nobody updates quietly decays, and because it keeps answering, the decay is easy to miss. Set a review rhythm, a named owner and an expiry date per version.

Frequently asked questions

What are Dhīr and EKA-1?

Dhīr is ZenteiQ's offline-capable, open-source model family (Dhīr 2B and Dhīr 5B) for document-grounded automation and on-premises multilingual assistance. EKA-1 is Soket Labs' large multilingual model, described by the company as a 26.3-billion-parameter open sparse hybrid model covering 22 scheduled Indian languages plus English.

Can my business use Dhīr or EKA-1 today?

That depends on availability, licensing and your own testing, none of which this article assumes. Confirm current access terms — the models are accessible via platforms such as AIKosh, per the launch report — and test on your own devices first. To discuss a pilot, contact Coheser.

Which workflow should I pilot first?

One frequent, document-based job your team already does on phones or tablets where connectivity is unreliable, and where mistakes are caught by staff before they reach customers. Expand only once the pilot beats the current method.