India’s IndiaAI Mission, a government programme under the Ministry of Electronics and IT, released three foundation models on 3 November 2026: EKA-1, developed by Soket Labs Technology and Research, and Dhīr 2B and Dhīr 5B, developed by ZenteiQ AiTech Innovations, The Hindu reported. IT Secretary S Krishnan attended the unveiling. The event took place in New Delhi, Swarajya reported. The release gives organisations building services for Indian languages new domestically developed models to work with, including models intended to operate without a continuous internet connection.
Key points
- Soket’s EKA-1 has 26.3 billion parameters and was trained on about 7.5 trillion tokens.
- EKA-1 covers all 22 scheduled Indian languages and English; ZenteiQ’s Dhīr models support English and 10 Indian languages.
- The government says subsidised national computing infrastructure and research-expense support enabled the models’ development.
- Both companies describe uses that can keep some AI processing on a local device.
IndiaAI’s subsidised computing programme
Both companies were selected under the IndiaAI Innovation Centre’s Foundation Models pillar, which supports the development of domestic AI capabilities and large language models. The government said access to subsidised national AI computing infrastructure, together with help for related research expenses, enabled development of the three models. That support concerns a costly part of model-building: the computing capacity used to turn training material into a working system.
The release brings two different kinds of model from that programme into view. Soket’s EKA-1 is intended for mathematics, code, reasoning and tasks in which a model takes actions towards a goal. ZenteiQ’s Dhīr models are described as smaller, multilingual systems suited to retrieving information from documents and assisting with services in places where connectivity is limited or information is sensitive. For companies building on them, the choice concerns both the languages a service needs to handle and where its processing would take place.
Krishnan said the models could strengthen India’s technological and strategic autonomy in AI and create opportunities to improve work across government, industry, research and education. Separately, AI Affairs previously reported India’s plan for an AI regulation consultation, with safety in focus. The model release concerns the infrastructure and systems being built under the mission.
Soket’s EKA-1 covers 22 Indian languages
EKA-1 has 26.3 billion parameters and was trained on about 7.5 trillion tokens. A parameter is a value adjusted during training, rather like the setting of a dial; a token is a piece of the text used in that process. Training changes those stored values as the model works through tokens, allowing it to generate responses later. The reported training total describes how much text was processed, while the parameter count describes the size of the model’s adjustable machinery.
The model covers all 22 scheduled Indian languages as well as English. Soket describes it as an open, sparse hybrid language model designed for mathematics, coding, reasoning and agentic use. Its chief executive, Abhishek Upperwal, said it was tailored to Indian banking, financial services and cybersecurity. Upperwal said Soket intends the model to meet the particular needs of those sectors and account for changes in India’s digital ecosystem.
EKA-1 can also perform certain inference on a local device such as a smartphone without mobile data or Wi-Fi, The Hindu reported. Inference is the work a trained model does when it processes a request and produces an answer. The local capability described for EKA-1 concerns certain inference, a narrower proposition than saying every use of the model runs on a phone.
Soket’s emphasis on language coverage and sector-specific use gives developers two practical considerations: whether the model handles the language needed for a service, and whether its behaviour suits the work being asked of it. The release describes the model’s training scale and intended applications, alongside that limited on-device capability.
ZenteiQ’s Dhīr models target local use
Dhīr 2B and Dhīr 5B are open-source models supporting English and 10 Indian languages. They are intended for document-grounded automation and information retrieval, including services in government, education, industry and defence. ZenteiQ’s chief executive, Prof. Sashikumar, said their lightweight design allows them to run on mobile phones without relying on the cloud, and that they can also run on graphics processing units and tensor processing units, chips used for AI calculations.
Keeping inference local changes where a request is handled: the device can process it instead of sending it to a remote service. The models are also described as deployable on premises as multilingual assistants, an option for settings where information is sensitive. Swarajya reported that Dhīr 2B and Dhīr 5B use a mixture-of-experts design and were trained from scratch using Google Cloud tensor processing units.
Finding an answer in a government document could therefore remain possible when an internet connection is unreliable, if one of the Dhīr models were deployed locally with the relevant document. The same arrangement could keep the request on the device. The models’ document-grounded and offline capabilities could make that arrangement possible in a low-connectivity setting.
For an organisation considering a deployment, the distinction between a model that can be run on premises and one that depends on a remote service would affect where its documents and requests are processed. ZenteiQ’s stated focus is on low-connectivity and sensitive environments, and Sashikumar said the models can operate across different computing platforms. Krishnan said platforms such as AIKosh could let people engaged in research, study or developing new ideas, as well as anyone else, experiment with the models.
