India unveils three indigenous AI models for domestic use

Launching domestic AI capabilities

India has officially introduced three indigenous artificial intelligence foundation models as part of the government's IndiaAI Mission. These models were unveiled in New Delhi during an event attended by IT secretary S Krishnan. The initiative represents a major move toward building the country's own domestic AI infrastructure.

The models were developed by two different startups that were selected under the IndiaAI Innovation Centre pillar. This specific pillar of the mission is designed to foster the creation of foundational models tailored to local needs. By supporting domestic development, the government aims to reduce reliance on external technologies.

The launch serves as a significant step in strengthening India's strategic autonomy in the field of artificial intelligence. According to S Krishnan, the initiative expands the availability of models that are specifically suited to domestic requirements. This includes applications in scientific research, coding, and mathematics.

Technical specs of the new models

Soket Labs introduced the EKA-1 model, which is a large-scale foundation model with 26.3 billion parameters. It was trained on nearly 7.5 trillion tokens to ensure high performance across various tasks. The model is designed to handle complex use cases such as reasoning, coding, and agentic workflows.

In contrast, ZenteiQ developed the Dhīr 2B and Dhīr 5B models, which are mixture-of-experts foundation models. These were trained from scratch using Google Cloud Tensor Processing Units, which are specialized hardware for accelerating machine learning tasks. These models are smaller and designed to function as multilingual assistants.

A key distinction between the models is their linguistic reach and hardware requirements. While EKA-1 supports all 22 scheduled Indian languages plus English, the Dhīr models support English and 10 other Indian languages. The Dhīr models are specifically optimized to run on mobile phones without requiring an internet connection.

Targeting specific industries

The developers of EKA-1 have focused on creating a tool that serves India's specific banking, financial services, and cybersecurity landscape. By tailoring the model to these sectors, the startup aims to provide high-utility AI for regulated industries. This focus helps address the unique security and operational needs of the Indian economy.

The Dhīr models offer a different utility by focusing on environments with limited connectivity. Because they can operate offline, they are useful in areas where stable internet access might be inconsistent. This makes them highly applicable for mobile users in diverse geographic locations across the country.

Furthermore, the models are capable of being deployed in air-gapped environments. An air-gapped environment is a computer system that is physically isolated from unsecured networks. This capability is particularly important for enterprise-based AI where data privacy and security are paramount requirements.

Open access and government support

The development of these models was made possible through subsidized AI compute and support for ancillary research expenses. These resources provided by the government helped the startups overcome the high costs associated with training large-scale foundation models. This support is a core component of the broader effort to strengthen the indigenous AI ecosystem.

Users can access these new models through the AIKosh platform, which serves researchers, enterprises, and the general public. This platform acts as a hub for distributing the tools and ensuring they reach the intended audiences. It facilitates the integration of these models into various commercial and academic projects.

ZenteiQ has taken a step toward open collaboration by releasing its Dhīr model weights under the Apache 2.0 license. This license allows others to use and modify the software freely. They also provided technical documentation and evaluation tools to help developers implement the models effectively.