New Delhi has just taken a big step towards building artificial intelligence that speaks India’s languages and works on India’s terms. On Tuesday, November 3, the Ministry of Electronics and Information Technology (MeitY) unveiled three homegrown AI foundation models, EKA-1, Dhir 2B and Dhir 5B, under the IndiaAI Mission, and the message from the launch was clear: India does not want to depend entirely on models built elsewhere for work that touches its banks, its government offices and its classrooms.
- What EKA-1 is built to do
- Why the Dhir models are different
- What this means for India’s AI push
- Frequently asked questions
- What are EKA-1, Dhir 2B and Dhir 5B?
- What is EKA-1 used for?
- Can the Dhir models work without the internet?
- Which languages do the models support?
- Where can researchers and developers access the models?
MeitY Secretary S. Krishnan unveiled the models in New Delhi with senior ministry and IndiaAI officials present. They come from two Indian builders, Bengaluru-based Soket Labs and ZenteiQ, an IISc Bangalore-born startup: one heavyweight built for complex reasoning, and two compact models built to run offline on ordinary laptops and mobile hardware.
What EKA-1 is built to do
EKA-1, developed by Soket Labs, is a 26-billion-parameter foundation model trained on about 7.5 trillion tokens. In plain terms, that scale puts it in the serious league for mathematics, coding and reasoning, the kind of work where a model has to follow a problem through several steps rather than just finish a sentence.
Soket Labs chief executive Abhishek Upperwal said the system was developed with India’s banking, financial services and cybersecurity requirements in mind. That focus matters because those are sectors where data cannot simply be shipped off to a foreign cloud for processing. A model that understands Indian languages and Indian use cases, and can be run in secure, controlled environments, gives banks and government agencies an option they did not have before.
EKA-1 is designed for Indian-language use, supporting all 22 languages listed in the Eighth Schedule of the Constitution alongside English.
Why the Dhir models are different
The Dhir models take the opposite route: small, open and offline. Developed by ZenteiQ, Dhir 2B and Dhir 5B are foundation models trained from scratch, rather than adapted from existing model weights, using Google Cloud Tensor Processing Units. They support English and 10 Indian languages and are built for document-grounded tasks, the kind of work where an answer has to come from a specific file, record or manual rather than from general knowledge.
The practical pitch is simple. Both models can run directly on tested mobile hardware and on laptops and edge devices without an internet connection. That makes them useful for field offices, clinics, classrooms and factory floors where connectivity is patchy, and for any enterprise that wants to fine-tune a model on its own data and run it on its own infrastructure.
- Dhir 2B: a dense model with 2.03 billion parameters, built for compact, offline deployment.
- Dhir 5B: a mixture-of-experts design with 5.09 billion total parameters, of which 1.79 billion are active at a time, balancing capability with efficiency.
- Open access: ZenteiQ is releasing the Dhir model weights under the Apache 2.0 licence, along with technical documentation and evaluation tools, through AIKosh.
- Who can use them: researchers, students, innovators, enterprises and the general public can access the models through AIKosh.
What this means for India’s AI push
The launch is one of the first public outputs from the startups the government picked to build sovereign models under the IndiaAI Mission’s foundation model pillar. Krishnan said indigenous models would strengthen India’s strategic autonomy in artificial intelligence and expand the availability of models suited to domestic requirements, supporting productivity across government, industry, research and education.
It is worth keeping the excitement in proportion. A launch is a milestone, not proof of large-scale commercial success, and independent benchmark results comparing the three models with established international systems are not yet available. What the launch does show is that Indian teams can now take a model from training to public release and put it in the hands of builders across the country.
The timing also fits a wider pattern. Indian firms are adopting AI fast but many are still working out how to measure returns and run AI safely at scale. Models that work offline, in Indian languages and inside an organisation’s own systems speak directly to that gap.
Frequently asked questions
What are EKA-1, Dhir 2B and Dhir 5B?
They are three indigenous AI foundation models unveiled by MeitY under the IndiaAI Mission on November 3, 2026. EKA-1 was built by Soket Labs, while Dhir 2B and Dhir 5B were built by ZenteiQ.
What is EKA-1 used for?
EKA-1 is a 26-billion-parameter model designed for mathematics, coding, reasoning and specialised enterprise work, with banking, financial services and cybersecurity among its target areas.
Can the Dhir models work without the internet?
Yes. Dhir 2B and Dhir 5B are designed to run offline on laptops, edge devices and tested mobile hardware, making them suitable for secure or low-connectivity settings.
Which languages do the models support?
EKA-1 supports all 22 Eighth Schedule Indian languages alongside English. The Dhir models support English and 10 Indian languages.
Where can researchers and developers access the models?
The models are accessible through AIKosh, with the Dhir model weights released under the Apache 2.0 licence along with documentation and evaluation tools.
India’s AI builders now have homegrown foundations to build on. The next test is not another launch event, but adoption: how many teams actually ship products on EKA-1 and Dhir in the months ahead.
