13 November 2026: India has added three new homegrown artificial intelligence models to its growing AI ecosystem. Under the government’s IndiaAI Mission, the Ministry of Electronics and Information Technology (MeitY) released EKA-1, developed by Soket Labs, and two compact models called Dhīr 2B and Dhīr 5B, developed by ZenteiQ AiTech Innovations, at an event in New Delhi on 3 November. Together, they cover very different needs, from powerful enterprise AI that understands all 22 scheduled Indian languages to small models that can run on a smartphone without internet. Here is a simple explanation of what these models are and why they matter.
What Was Launched?
The models were unveiled in the presence of MeitY Secretary S. Krishnan and senior IndiaAI officials. Both developers are based in Bengaluru, and the models were built under the IndiaAI Innovation Centre’s foundation models initiative, which supports Indian companies in creating AI systems trained for Indian needs. The new releases add to earlier models backed by the mission, including those from Sarvam AI, BharatGen and Gnani AI.
EKA-1: The Big Multilingual Model
EKA-1 is the larger of the new models. According to official information and multiple reports:
- It has 26.3 billion parameters, which broadly reflects the size and capacity of the model.
- It was trained on roughly 7.5 trillion tokens of text.
- It supports all 22 scheduled Indian languages as well as English.
- It is described as an open, sparse hybrid model designed for mathematics, coding, reasoning and agentic tasks.
Soket Labs has said the model was developed with sectors such as banking, financial services and cybersecurity in mind, while reports also list government, defence, education and industry as target uses. A sparse model activates only part of its network for each task, which can make it more efficient to run than a dense model of the same size.
Dhīr 2B and Dhīr 5B: Small Models for Phones
The Dhīr models take a very different approach. With about 2 billion and 5 billion parameters respectively, they are compact enough to run directly on smartphones and other everyday devices. Key features include:
- Offline use: They can work without an internet connection, which helps in areas with poor connectivity and in secure environments.
- Indian languages: They support English and 10 Indian languages.
- Open source: Developers can download, study and build on them.
Small, offline models are useful for things like local translation, voice assistants, document help and field work, where sending data to the cloud is slow, costly or not allowed.
EKA-1 vs Dhīr: Quick Comparison
| Feature | EKA-1 | Dhīr 2B / 5B |
|---|---|---|
| Developer | Soket Labs | ZenteiQ AiTech Innovations |
| Size | 26.3 billion parameters | 2 billion and 5 billion parameters |
| Languages | 22 scheduled Indian languages and English | English and 10 Indian languages |
| Main strengths | Reasoning, maths, coding, agentic tasks | Lightweight, offline use on devices |
| Typical users | Enterprises, government, researchers | App developers, field and secure deployments |
Where Can You Access Them?
According to MeitY, the models are being made available to researchers, students and enterprises through AIKosh, the IndiaAI Mission’s platform for datasets, models and AI tools. Developers interested in building products on top of them can explore the listings and documentation there.
Why This Matters for India
Strategic Autonomy
Speaking at the launch, S. Krishnan said indigenous foundation models strengthen India’s technological and strategic autonomy in AI. Relying only on foreign models raises questions about data control, cost and access in sensitive sectors.
Language Inclusion
Most popular AI tools work best in English. Models trained deeply on Indian languages can help millions of people use AI in their own language, whether for government services, education, farming advice or banking.
Cost and Local Deployment
Smaller and sparse models can be cheaper to run and can be deployed within India or even on devices, which matters for privacy and for organisations that cannot send data abroad.
Challenges Ahead
Launching a model is only the first step. Experts will be watching how well these models perform on real-world tasks compared with global alternatives, whether Indian startups and enterprises actually adopt them, and whether there is a steady plan for compute and data to keep updating them. Independent benchmarks and developer feedback over the coming months will give a clearer picture.
What to Watch Next
India’s AI calendar remains busy. The Bengaluru Tech Summit, scheduled for 17 to 19 November 2026, is expected to feature discussions on AI, semiconductors and startups, and more announcements under the IndiaAI Mission are likely in the months ahead.
Frequently Asked Questions
Who developed EKA-1?
EKA-1 was developed by Bengaluru-based Soket Labs Technology and Research under the IndiaAI Mission.
Can Dhīr models work without the internet?
Yes. Dhīr 2B and Dhīr 5B are designed to be small enough to run on smartphones and other devices offline.
How many languages does EKA-1 support?
It supports all 22 scheduled Indian languages along with English.
Are these models free to use?
The Dhīr models are open source. Access terms for each model are listed on the AIKosh platform.
Dedicated to researching and publishing informative content related to gig work, freelance opportunities, remote jobs, online earning options, and career development. Committed to delivering accurate, reader-friendly, and useful information for individuals seeking flexible work opportunities.