India launches three indigenous AI models for secure offline use

A push for strategic autonomy

The Ministry of Electronics and Information Technology (MeitY) has officially unveiled 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 government intends for these tools to enhance India's strategic autonomy in the global AI landscape.

By developing these models locally, the government aims to provide options that are specifically suited to domestic requirements. This initiative moves away from a total reliance on international providers for core AI infrastructure. It creates a foundation for Indian companies to build specialized applications that align with local needs.

MeitY Secretary S. Krishnan highlighted that the initiative is about more than just general-purpose technology. He noted that the models are designed to strengthen the country's position in the AI space. The goal is to ensure that India has the resources necessary to develop its own technological capabilities independently.

Technical specs and capabilities

The EKA-1 model was developed by a Bengaluru-based company called Socket AI. It is a 26-billion-parameter foundation model specifically designed for complex tasks like mathematics, coding, and reasoning. To achieve this, the model was trained on 7.5 trillion tokens and supports over 100 different languages, including many Indian and international languages.

In contrast, Dhīr 2B and Dhīr 5B were developed by ZenteiQ as mixture-of-experts foundation models. These were trained from scratch using Google Cloud Tensor Processing Units (TPUs). These models are smaller and more efficient, supporting English and 10 Indian languages while being capable of running directly on tested mobile hardware.

A key technical feature of these models is their ability to operate in private environments. They are designed to function in on-premises setups, private clouds, and air-gapped environments. An air-gapped environment is a computer system that is physically isolated from unsecured networks, such as the public internet, to prevent unauthorized access.

Enterprise and research use

The government is focusing on specific applications for these models rather than just general use. These include fields such as scientific research, cybersecurity, and advanced coding. Because the models can run in air-gapped environments, they are particularly useful for industries that handle highly sensitive data.

S. Krishnan stated that these models could make enterprise-based artificial intelligence possible in areas where it hasn't taken off yet. Many companies are hesitant to use cloud-based AI because of security risks. These indigenous models provide a way to use AI while keeping all data within a company's own secure perimeter.

The models are also intended to improve productivity and efficiency across various different applications. Smaller models like the Dhīr series are particularly useful for these types of tasks. The government plans to continue supporting the refinement and development of these models to ensure they stay effective as technology evolves.

Accessibility and licensing

ZenteiQ has decided to release the model weights for Dhīr 2B and Dhīr 5B under the Apache 2.0 license. Model weights are the learned parameters that define how an AI model processes information. By releasing these, the company allows others to use and build upon the models more freely.

The release includes not just the weights, but also technical documentation and evaluation tools. These resources are being made available through a platform called AIKosh. This platform serves as a hub for researchers, students, innovators, and the general public to access these national resources.

By making these models accessible to a wide audience, the government hopes to foster a broader ecosystem of innovation. Students and researchers can use the models to learn and experiment with the underlying technology. Meanwhile, enterprises can use them to build secure, specialized tools for their specific business needs.