IndiaAI Mission has released three indigenous foundation models – EKA-1, Dhīr 2B and Dhīr 5B. The launch is important because these models have been developed in India with a focus on Indian languages, local requirements and secure deployment.

The models are not consumer apps in the same way as a chatbot website. They are the technology layer on which developers, government departments, schools, banks and businesses can build their own AI tools.

EKA-1 was developed by Soket Labs Technology and Research, while Dhīr 2B and Dhīr 5B were developed by ZenteiQ AiTech Innovations. The projects received support under the IndiaAI Innovation Centre pillar of the IndiaAI Mission reported PIB.

What is a foundation model

A foundation model is an AI system trained on a very large collection of text, code and other information patterns. It learns how language works, how questions are usually answered and how ideas relate to each other.

Developers can then adapt the model for specific tasks. For example, a bank may use a model to help employees find answers in policy documents, while a school may use it to create study support in an Indian language.

A foundation model does not simply store a ready-made answer for every question. It predicts useful words, numbers or code based on patterns learned during training. This is why AI can write, summarize, translate and answer questions. It is also why important outputs still need human review.

EKA-1 – the larger multilingual model

EKA-1 is Soket Labs’ open, sparse hybrid language model. It has 26.3 billion parameters and was trained on about 7.5 trillion tokens. Parameters are the internal settings an AI model uses to recognize patterns, while tokens are small units of text that the model processes during training.

According to the IndiaAI announcement, EKA-1 supports all 22 scheduled Indian languages along with English. It is designed for mathematics, coding, reasoning and agent-based tasks, where AI can complete steps within defined instructions.

Soket Labs has positioned EKA-1 for areas such as banking, financial services and cybersecurity. A practical example could be an internal banking assistant that helps staff search policy documents in Hindi, Tamil or English without moving sensitive information to a public AI service.

Dhīr 2B and Dhīr 5B – smaller models for secure use

ZenteiQ’s Dhīr 2B and Dhīr 5B are open-source multilingual models. The “2B” and “5B” refer to their approximate model size in billions of parameters.

They support English and 10 Indian languages and are designed for on-premise and offline-capable use. On-premise means the model can run inside an organization’s own computer systems instead of sending sensitive documents to an outside cloud service as per Newsonair.

This can be useful in sectors such as defence, government, education and industry. For example, a district office could use a Dhīr-based tool to search circulars and answer routine citizen queries even where internet connectivity is weak. An organization handling confidential documents could keep data within its own network.

Smaller models may not match a large cloud-based model on every complex task. Their advantage is that they can be more affordable to run, faster for focused tasks and easier to use in controlled environments.

Why these Indian AI models matter

The biggest benefit is language relevance. Many global AI systems work best in English and may not understand Indian languages, local terms or administrative context equally well. Models trained with Indian-language support can make digital services more useful for a wider population.

The launch also supports data control. Organizations in sensitive sectors may prefer AI systems that can run locally, especially when their documents include financial, personal or government information.

Another benefit is access for researchers and startups. IndiaAI has said that the models can be made available through AIKosh, helping students, innovators and developers experiment with Indian AI tools.

How they compare with global AI models

EKA-1 and Dhīr will compete indirectly with international models from companies such as OpenAI, Google, Meta and Anthropic. But the goal is not only to build the largest model.

Their stronger focus is on Indian languages, local deployment and sector-specific needs. In many cases, a smaller model that works reliably with a company’s own documents may be more practical than a much larger public AI model.