India releases three indigenous AI models for local and multilingual use

Launch of indigenous models

The IndiaAI Mission recently unveiled three new indigenous foundation models during an event in New Delhi on November 3. These models were developed by two different companies, Soket AI Labs and ZenteiQ AiTech Innovations, under the IndiaAI Innovation Centre’s Foundation Models pillar. The launch marks a significant step toward creating AI tools that are specifically designed for the Indian context.

The models were introduced in the presence of S Krishnan, who serves as the Secretary at the Ministry of Electronics and Information Technology (MeitY). He noted that developing these indigenous models would help strengthen India's technological autonomy in the field of artificial intelligence. This initiative is intended to create new opportunities to improve productivity across various sectors, including research, education, and government.

Sudeep Shrivastava, the COO of IndiaAI, explained that these releases show progress across different categories of AI development. The new models range from lightweight systems designed for everyday devices to more complex models built for enterprise applications. By providing these tools, the mission aims to support a wide variety of use cases for both individual users and large organizations.

EKA-1 for finance and security

One of the primary models released is EKA-1, which was developed by Soket AI Labs. This model is described as an open, sparse hybrid language model with 26.3-billion parameters. It was trained on approximately 7.5 trillion tokens to ensure it can handle complex tasks like reasoning, coding, and mathematics.

Soket AI Labs designed EKA-1 specifically to meet the requirements of India's banking, financial services, and cybersecurity sectors. The company stated that the model is built to understand India's evolving digital ecosystem and deliver solutions that align with the country's specific priorities. It is also capable of supporting all 22 scheduled Indian languages in addition to English.

Abhishek Upperwal, the CEO of Soket AI Labs, said the model was tailored to India's unique banking, financial services and cybersecurity landscape. By focusing on these specific industries, the developers hope to provide high-quality AI support for critical infrastructure. The model's ability to handle agentic applications also makes it useful for automated workflows in these sensitive environments.

Offline and mobile AI

The second set of models, Dhīr 2B and Dhīr 5B, was introduced by ZenteiQ AiTech Innovations. These are open-source multilingual assistants that support English and 10 Indian languages. Unlike some larger models that require constant internet access, these were built for on-premise and offline deployment.

The Dhīr models are designed for specific tasks such as document-based automation, information retrieval, and providing services to citizens and enterprises. Because they can operate without continuous cloud connectivity, they are particularly useful for government and defense sectors where data security is a high priority. They can be deployed across various computing platforms, including GPUs and TPUs.

Prof Sashikumar, the CEO of ZenteiQ AiTech Innovations, highlighted that the lightweight architecture of the Dhīr models allows them to run on mobile phones. He explained that this allows users to access AI capabilities locally while keeping data on their devices. This approach is intended to enhance privacy, security, and accessibility for people who may not have constant access to high-speed internet.

Infrastructure and access

The development of these three models was supported by the IndiaAI Mission through the IndiaAI Innovation Centre. The mission provided the participating teams with access to subsidized national AI compute infrastructure. They also provided support for the research expenses required to train and refine these foundation models.

By offering subsidized compute, the mission helps reduce the high costs typically associated with training large-scale AI models. This allows smaller companies and research teams to compete in the development of high-quality technology. The goal is to foster a local ecosystem where multiple models can be built for different regional needs.

Researchers, students, and developers can access these models through a platform called AIKosh. This platform is part of the IndiaAI Mission and is designed to facilitate experimentation. By making the models available on AIKosh, the government hopes to encourage further innovation and the creation of new applications based on these indigenous foundations.