India has added three more homegrown artificial intelligence models to its growing list. The Ministry of Electronics and Information Technology (MeitY) on Tuesday launched EKA-1, built by Soket AI, and the Dhīr 2B and Dhīr 5B models, built by ZenteiQ, under the Foundation Model initiative of the IndiaAI Mission at an event in New Delhi.
The pitch is simple: Indian-made AI that can run on private servers, in sensitive sectors where data cannot leave the building, and even on an ordinary mobile phone without an internet connection.
EKA-1: A 26-billion-parameter model for enterprises
EKA-1 is the larger of the new models. Developed by Soket AI, it has about 26 billion parameters and was trained on roughly 7.5 trillion tokens of data. The company describes it as an open, sparse hybrid language model designed for mathematics, coding, reasoning and AI agents.
Its focus is on enterprise and high-security use. EKA-1 can be deployed on private servers, private clouds and in environments that have no internet access at all, which makes it suitable for sectors such as banking, financial services and cybersecurity. Soket AI says all its data preparation, training and evaluation were done in India.
"We want to remove the reliance of foreign models," Soket AI CEO Abhishek Upperwal said at the launch. The company plans to release EKA-1's weights openly within about a month, with early access through a waitlist.
Dhīr 2B and 5B: AI that lives on your phone
The two Dhīr models, from ZenteiQ, are much smaller, at around 2 billion and 5 billion parameters. That small size is the whole point. They are designed to run directly on a mobile phone without an internet connection, which could bring AI tools to villages and remote areas with poor connectivity.
The models support English and several Indian languages and are aimed at tasks like document search, information retrieval and task automation. Possible users include government departments, defence, education and industries where privacy is critical.
The Dhīr model weights are already available on AIKosh, the government's AI datasets and models platform. ZenteiQ is also building a free app for Android and iOS. Founder Sashikumaar Ganesan explained that once downloaded, the model "will be sitting in the user's mobile itself". The company also plans a larger 9-billion-parameter model in December for researchers and engineers.
Quick comparison
| Feature | Details |
|---|---|
| Launched by | MeitY, under the IndiaAI Mission's Foundation Model initiative |
| EKA-1 developer | Soket AI |
| EKA-1 size | About 26 billion parameters, trained on about 7.5 trillion tokens |
| EKA-1 best for | Maths, coding, reasoning, AI agents; banking, finance, cybersecurity |
| EKA-1 availability | Open weights planned within a month; waitlist access first |
| Dhīr developer | ZenteiQ |
| Dhīr sizes | 2B and 5B parameters |
| Dhīr best for | Offline use on phones, document search, task automation |
| Dhīr availability | Weights on AIKosh; free Android and iOS app in development |
Why this matters for India
India is one of the world's biggest markets for AI tools, but most of the popular models are built abroad and run on foreign cloud servers. That raises concerns about data privacy, cost and strategic dependence, especially for government, defence and banking.
Models like EKA-1 and Dhīr try to fill that gap. MeitY Secretary S Krishnan said the models could improve productivity across government, businesses, research and education, and highlighted their value for "air-gapped" systems that have no connection to outside networks.
The offline phone angle could be the bigger story for ordinary users. A farmer, student or small shopkeeper in an area with patchy mobile data could, in theory, use an AI assistant that works entirely on the handset. How well these compact models perform in real-world Indian language use will decide whether that promise holds up.
How the IndiaAI Mission helped
Under the IndiaAI Mission, selected startups receive subsidised access to computing power and funding for research costs to build foundation models. These latest launches add to the earlier models released under the same programme, as the government pushes for a homegrown AI stack.
What happens next
Developers and researchers can already download the Dhīr weights from AIKosh, while EKA-1's open release is expected over the coming weeks. Watch for ZenteiQ's mobile app and its 9-billion-parameter model in December, along with independent benchmarks that will show how these Indian models compare with global open-source rivals.
Comments (0)