IndiaAI Mission Released Indigenous Foundation Models EKA-1 by Soket and Dhīr 2B & Dhīr 5B by ZenteiQ Under the IndiaAI Innovation Centre Pillar

EKA-1 and Dhīr 2B and Dhīr 5B are large AI models developed by Soket Labs Technology and Research Private Limited and ZenteiQ AiTech Innovations. They are designed to handle language and other complex tasks. Their importance lies in providing home-grown AI capabilities tailored to India’s languages, institutions, and sectors.

EKA-1 supports all 22 scheduled Indian languages and English. It is designed for mathematics, coding, reasoning, and agentic use. Dhīr models support English and 10 Indian languages. They focus on multilingual assistance, document-grounded automation, knowledge retrieval, and citizen or enterprise services.

They are called foundation models because they can serve as general bases for many downstream uses, rather than performing only one narrow task. The article presents them as resources for government, industry, research, education, defence, and other critical sectors. Their open availability can also encourage experimentation through platforms such as AIKosh.

EKA-1 has 26.3 billion parameters and was trained on approximately 7.5 trillion tokens. It covers all 22 scheduled Indian languages alongside English. The article does not give parameter counts or training-token totals for Dhīr 2B and Dhīr 5B, but their names identify sizes of 2 billion and 5 billion parameters.

Dhīr 2B and Dhīr 5B support English and 10 Indian languages. This makes EKA-1 broader in stated language coverage. EKA-1 is also described as a sparse hybrid language model, built for mathematics, code, reasoning, and agentic use.

The figures show two different approaches to scale. EKA-1 combines a very large parameter count and extensive training data with wide language coverage. Dhīr offers smaller models intended for practical, secure deployment. The article emphasizes that Dhīr can operate on-premise and offline, including in low-connectivity settings.

EKA-1 is Soket’s larger, open, sparse hybrid language model. It has 26.3 billion parameters and was trained on approximately 7.5 trillion tokens. Its stated strengths include mathematics, code, reasoning, and agentic use across all 22 scheduled Indian languages and English.

Dhīr 2B and Dhīr 5B are ZenteiQ’s smaller open-source models. They support English and 10 Indian languages. Their main functions include document-grounded automation, knowledge retrieval, and citizen or enterprise services. They are designed as multilingual assistants rather than as the article’s broad, high-capability model.

Deployment is the clearest contrast. EKA-1 is presented as a powerful foundation model, including applications for banking, financial services, and cybersecurity. Dhīr is explicitly designed for secure, on-premise deployment and offline-capable operation across defence, government, industrial, and educational settings, especially where connectivity is limited.

Open-source Dhīr models are made available for organizations to use and adapt under their applicable open-source terms. On-premise deployment means the models can run on an organization’s own computing systems instead of relying entirely on an external service. This can give government, defence, education, and industry greater control over deployment.

Offline-capable operation means Dhīr can support tasks even when a reliable internet connection is unavailable. For example, an institution could use Dhīr to search internal documents, automate document-based work, or provide services in supported languages. The key mechanism is document grounding: the assistant uses relevant organizational information when responding.

The article specifically highlights secure, on-premise, multilingual assistants for critical sectors and low-connectivity environments. It does not provide detailed security testing or deployment requirements. Still, local operation could reduce dependence on continuous connectivity and make deployment more practical where sensitive work or limited networks are involved.

Soket and ZenteiQ were selected under the IndiaAI Innovation Centre’s Foundation Models pillar. The IndiaAI Mission, under MeitY, gave them access to subsidized national AI compute infrastructure. It also supported ancillary research expenses connected with developing their models.

This support matters because training advanced AI requires substantial computing resources, data processing, experimentation, and engineering. In the article’s example, Soket used this ecosystem to develop EKA-1, while ZenteiQ introduced the Dhīr 2B and Dhīr 5B models. Public infrastructure helped both teams pursue indigenous foundation AI capabilities.

The broader significance is that national computing can accelerate home-grown innovation. It can help Indian teams build models suited to local languages and needs. The article describes these models as strengthening technological and strategic autonomy, while also creating resources for researchers, students, innovators, and the public through platforms such as AIKosh.

Locally developed models could make AI more useful across India’s linguistic and institutional landscape. EKA-1 supports all 22 scheduled Indian languages and English, while Dhīr supports English and 10 Indian languages. That coverage could broaden access to automation, information services, education, research, and government support.

The models also offer different practical routes. EKA-1 is designed for mathematics, coding, reasoning, and agentic use, including India’s banking, financial services, and cybersecurity landscape. Dhīr can provide document-grounded automation and retrieval on-premise, with offline-capable operation for defence, government, industry, and education.

The article presents these developments as opportunities to improve productivity, efficiency, experimentation, and strategic autonomy. It does not claim that every organization will immediately adopt them or that all risks are solved. Their impact will depend on deployment, evaluation, and continued innovation, but domestic models create a stronger base for India-specific AI services.

In general, a language model learns by processing tokens, which are pieces of text such as words or word fragments. During training, it studies relationships among tokens and adjusts internal parameters to predict useful continuations. EKA-1 was trained on approximately 7.5 trillion tokens and has 26.3 billion parameters.

When a user provides a prompt, the model converts it into tokens and uses its learned parameters to estimate what should come next. Repeated predictions can produce an answer, computer code, a reasoning response, or an agent-like sequence of actions. Dhīr’s document-grounded design can also connect responses to supplied documents, supporting retrieval and automation.

These mechanics explain how one model can support many tasks, but they do not guarantee accuracy. The article reports intended capabilities, not independent performance results. EKA-1 is aimed at mathematics, code, reasoning, and agentic use, while Dhīr focuses on multilingual assistance and document-based services.