India unveils three indigenous AI models for offline and secure use

A push for strategic autonomy

The Ministry of Electronics and Information Technology (MeitY) recently introduced three new artificial intelligence foundation models. These models, named EKA-1, Dhīr 2B, and Dhīr 5B, were developed by domestic companies with support from the IndiaAI Mission. The initiative aims to build a stronger foundation for the country's own AI capabilities.

MeitY Secretary S. Krishnan noted that the project is intended to strengthen India's strategic autonomy in the field of artificial intelligence. By creating models that meet specific domestic requirements, the government hopes to reduce reliance on external technologies. This move is part of a broader effort to make AI more accessible and relevant to the local landscape.

The models were created by two specific companies based in Bengaluru. Socket AI developed the EKA-1 model, while ZenteiQ was responsible for the Dhīr 2B and Dhīr 5B models. Both companies worked under the Foundation Model initiative to ensure the technology serves a wide range of national interests.

Technical specs and capabilities

The EKA-1 model is a 26-billion-parameter foundation model specifically designed for complex tasks like mathematics, coding, and reasoning. It was trained on 7.5 trillion tokens and is capable of supporting more than 100 different languages. This includes a variety of Indian and international languages, as well as various programming languages.

In contrast, the Dhīr 2B and Dhīr 5B models use a mixture-of-experts architecture. This means they are designed to be more efficient by activating only the most relevant parts of the model for a given task. These models were trained from scratch using Google Cloud Tensor Processing Units (TPUs) and support English along with 10 Indian languages.

A key technical distinction is the hardware compatibility of the Dhīr models. They are designed to run directly on tested mobile hardware without requiring an active internet connection. This makes them highly portable and useful for applications where constant connectivity is not possible or desired.

Security and offline deployment

One of the most significant features of these models is their ability to operate in air-gapped environments. An air-gapped environment is a secure network that is physically isolated from the public internet. This capability is particularly important for organizations that handle highly sensitive information.

S. Krishnan highlighted that these models could enable enterprise-based artificial intelligence to take off in India. Many companies have been hesitant to adopt AI because of the risks associated with sending private data to external clouds. These models allow for on-premises, private cloud, and air-gapped deployment to solve that problem.

The models are intended for specialized applications rather than just general-purpose use. They are expected to be useful in fields such as scientific research, cybersecurity, and advanced mathematics. By providing a secure way to run these tools, the government hopes to see wider adoption across various industrial sectors.

Open access and availability

ZenteiQ has decided to release the model weights for the Dhīr 2B and Dhīr 5B models under the Apache 2.0 license. This is an open-source license that allows others to use, modify, and distribute the software. They are also providing technical documentation and evaluation tools to help users get started.

All three models are being made available through a platform called AIKosh. This platform serves as a national resource for a variety of different users. It provides access to researchers, students, innovators, and the general public, as well as various enterprises.

The government plans to continue supporting the refinement and development of these models over time. By making them accessible to a wide audience, the goal is to foster innovation across the country. The initiative seeks to provide the building blocks for others to create specialized tools and applications.

Impact on the AI landscape

The launch of these models represents a shift toward localized and specialized AI infrastructure. Instead of relying solely on massive, general-purpose models from global tech giants, India is building tools tailored for specific needs. These include smaller models that can improve productivity and efficiency in niche applications.

For businesses, the ability to run AI locally means they can maintain better control over their data privacy. This could lead to a surge in AI adoption within government agencies and private companies that deal with restricted data. It levels the playing field for organizations that cannot afford to move their entire infrastructure to the cloud.

Looking ahead, the success of these models will depend on how well they are integrated into existing workflows. The government's continued support for their development will be key to ensuring they remain competitive. As these models become more refined, they may become standard tools for Indian researchers and developers.