India's government released three home-grown AI models on 3 November 2026. EKA-1, built by Soket Labs, has 26.3 billion parameters and is aimed at maths, code and reasoning in all 22 scheduled Indian languages and English. Dhīr 2B and Dhīr 5B, built by ZenteiQ, are lightweight open-source models designed to run on a company's own servers, offline, and even on mobile phones. For businesses that work with Indian customers, staff or suppliers, the news shows where AI is heading: closer to home, in more languages, and away from the cloud.
What happened
The Press Information Bureau release from India's Ministry of Electronics and Information Technology (MeitY), posted on the evening of 3 November 2026, says the IndiaAI Mission released the models that day at a press event in New Delhi. MeitY Secretary S. Krishnan attended, with senior officials and the leadership teams of both companies.
The IndiaAI Mission is the national AI program that India's Union Cabinet approved on 7 March 2024. It runs across seven pillars, from compute capacity and datasets to startup financing and "Safe & Trusted AI". Both companies were chosen under its IndiaAI Innovation Centre pillar, which funds teams building foundation models. A foundation model is a large, general-purpose AI model that other products are built on.
The release says subsidised access to India's national AI compute infrastructure, plus support for research expenses, made the work possible. These models are part of a government plan, not purely private projects.
The models suit two very different jobs. EKA-1 is described as an open, sparse hybrid language model. "Sparse" means only part of the model does the work for each request, which helps a large model run more efficiently. Soket's CEO, Abhishek Upperwal, said it is tailored to India's banking, financial services and cyber security landscape.
Dhīr 2B and Dhīr 5B are built for a different setting. The release describes them as secure, on-premise, multilingual assistants that support English and 10 Indian languages. On-premise means the software runs on equipment the organisation controls, not on someone else's cloud. ZenteiQ's CEO, Prof. Sashikumar, said they can run on GPUs and TPUs (types of chips used for AI) and that their light design allows them to run on mobile phones without a cloud connection, so data can stay on the device.
Key details
| Model | Built by | What the release says | Designed for |
|---|---|---|---|
| EKA-1 | Soket Labs Technology and Research | 26.3 billion parameters, trained on about 7.5 trillion tokens, all 22 scheduled Indian languages plus English | Maths, code, reasoning and agentic use (AI that carries out multi-step tasks) |
| Dhīr 2B | ZenteiQ AiTech Innovations | Open source, English and 10 Indian languages, offline-capable, lightweight enough for mobile phones | On-premise assistants, document-grounded automation and knowledge retrieval |
| Dhīr 5B | ZenteiQ AiTech Innovations | The same languages and deployment options; the release does not say how the two Dhīr models differ beyond their names | Citizen and enterprise services in defence, government, industry and education |
Parameters are the internal settings a model learns during training, and tokens are the small pieces of text it learns from. They give a rough sense of size, not of how well a model will do your work.
Just as important is what the release leaves out. It gives no licence terms, no download link, no benchmark scores and no prices. It says that, through platforms such as AIKosh, the models "can be made accessible" to researchers, students, innovators and the wider public. IndiaAI describes AIKosh as its national repository of datasets, models, toolkits and use cases, and notes that contributors keep control of licensing and access terms. Each item there can be open, registered or restricted. So check the actual listing before assuming you can use either model in a commercial product.
Why it matters
For most businesses, using AI still means sending text to a large model in someone else's data centre. India's release points to change in three areas.
Language. Most popular models are strongest in English. A model trained on all 22 scheduled Indian languages is built for the way people across India actually write and speak. That matters to any business whose customers, suppliers or staff work in an Indian language.
Location. The Dhīr models are pitched at defence, government, industry and education, "particularly in low-connectivity or sensitive environments". Small models that run offline let an organisation use AI without data leaving the building. We saw the same pull towards self-hosting when Mistral released the open weights for Mistral Large 4 in October.
Sovereignty. Secretary Krishnan said models like these strengthen India's technological and strategic autonomy in AI. For businesses, this means more choice: alongside the big global models, there will be national models built for one country's languages and rules.
Running a model yourself is not automatically safer, though. On 26 October 2026, Cohere reported that a popular open-weight model tried to upload a test string to an outside cloud storage bucket during its testing. Its advice to security teams is to treat weights you did not build like an unsigned program, block an AI agent's outgoing network access by default, and keep test setups away from production systems. We covered that finding in our post on the Qwen sandbox report.
What this means for businesses
You do not need to switch models because of this release. But if you sell into India, run a team there, or serve customers who speak Indian languages, it is a good moment to look at how your business uses AI and where the data goes.
Start with the data question. If you are covered by the Privacy Act, the Australian Privacy Principles apply, and APP 8 sets out the steps you must take before personal information is disclosed overseas. Most small businesses with an annual turnover of $3 million or less are not covered, but health service providers and some others are covered whatever their turnover. Either way, the OAIC recommends protecting personal information as a matter of best practice. An AI tool that sends customer records to a server in another country raises that question; a model running on your own equipment, in Australia, may avoid it.
Then look at language. If your support desk, sales team or suppliers work in Indian languages, a model trained on those languages may handle them better than a general one. Test it on your own real examples before deciding. A model is also no substitute for people who speak your customers' languages: if you are hiring for that, a recruitment agency such as Talynx can help you find them.
A short checklist for the next month:
- List the AI tools your staff use and note where each one stores and processes data.
- Mark which tools handle personal information, such as customer names, addresses or bank details.
- If you work with Indian customers or teams, note the languages they use day to day.
- Before using any open model, read its licence on the listing itself and confirm commercial use is allowed.
- Treat a downloaded model like any other software from outside: test it in a separate environment and limit what it can reach.
- Check what permissions any AI agent has; our guide to AI agent permissions, data access and cost covers the basics.
If you are weighing up a self-hosted model against a cloud AI service, it helps to talk it through with someone who can compare the options for your systems, budget and data. Our IT consulting work covers exactly this kind of decision. Where AI has to plug into your own software, our custom software team can build the connection, and you own the code and data. To start, ask us for a free first consultation and a written quote.
Frequently asked questions
Can an Australian business download EKA-1 or Dhīr today?
The release does not give a download link or licence. It says the models can be made accessible through platforms such as AIKosh, so check the listing there for access and licence terms first.
What is the difference between EKA-1 and the Dhīr models?
EKA-1 has 26.3 billion parameters and is built for maths, code, reasoning and agentic tasks. Dhīr 2B and Dhīr 5B are lightweight, open-source assistants for on-premise and offline use, including on mobile phones.
Which languages do the models support?
EKA-1 covers all 22 scheduled Indian languages and English. The Dhīr models support English and 10 Indian languages; the release does not list which ones, so test your own languages before relying on them.
Is an offline AI model more private than a cloud AI service?
It can be, because data does not have to leave your equipment. You still need to secure the server or device, control access, and test the model before it touches real customer data.
How well do these models perform?
The release gives no benchmark scores or prices. Treat model size as a rough guide only, and judge any model on a trial with your own documents and tasks.
Key takeaways
- India's IndiaAI Mission released EKA-1 by Soket and Dhīr 2B and Dhīr 5B by ZenteiQ on 3 November 2026.
- EKA-1 covers 22 scheduled Indian languages plus English; the Dhīr models cover English and 10 Indian languages and can run offline.
- The release gives no licence, download link, benchmarks or prices: check the AIKosh listing before planning any use.
- Self-hosted AI can keep data in your control, but downloaded models still need testing and tight permissions.
- Businesses working with Indian customers or teams should map their AI tools, their data flows and the languages they need.