India Goes Local: MeitY Unveils Indigenous AI Models EKA-1 and Dhīr
India launches EKA-1, Dhīr 2B, and Dhīr 5B indigenous AI models for coding, math, and offline use to boost strategic autonomy in technology.
In a massive leap toward technological self-reliance, the Ministry of Electronics and Information Technology (MeitY) has officially unveiled three homegrown artificial intelligence foundation models: EKA-1, Dhīr 2B, and Dhīr 5B. Developed by domestic firms under the ambitious IndiaAI Mission, these models are designed to reduce dependency on foreign technology and provide secure, specialized AI capabilities tailored to India's unique needs.
MeitY Secretary S. Krishnan emphasized that this initiative is a cornerstone for strengthening India's strategic autonomy in the AI landscape. By fostering the creation of models that meet domestic requirements, the government aims to expand the reach of AI across various sectors while ensuring data security and sovereignty.
EKA-1: The Powerhouse for Coding and Logic
Developed by the Bengaluru-based firm Socket AI, EKA-1 is a heavyweight in the indigenous AI space. With a staggering 26-billion-parameter architecture, this foundation model is specifically engineered to excel in complex domains such as mathematics, coding, and logical reasoning.
The technical depth of EKA-1 is impressive, having been trained on 7.5 trillion tokens. Its versatility is evident in its linguistic capabilities, supporting over 100 international and Indian languages, alongside various programming languages. Beyond simple queries, EKA-1 is built to:
- Handle intricate, multi-step tasks.
- Execute code directly.
- Operate within private clouds, on-premises setups, and air-gapped environments.
Dhīr 2B and 5B: AI That Works Offline
While EKA-1 focuses on power and reasoning, the Dhīr 2B and Dhīr 5B models, developed by ZenteiQ, focus on accessibility and efficiency. These are "mixture-of-experts" foundation models trained from the ground up using Google Cloud Tensor Processing Units (TPUs).
One of the most groundbreaking features of the Dhīr series is its ability to run directly on tested mobile hardware without requiring an internet connection. Supporting English and 10 Indian languages, these models bring the power of AI to the palm of the hand, regardless of connectivity.
In a move to support the open-source community, ZenteiQ is releasing the model weights under the Apache 2.0 licence. Technical documentation and evaluation tools are being made available through AIKosh, ensuring that the developer community can build upon this foundation.
Breaking the Barrier: Air-Gapped AI and Enterprise Growth
A critical highlight of this launch is the capability of these models to function in air-gapped environments—systems that are physically isolated from unsecured networks. Secretary S. Krishnan noted that this feature is a game-changer for enterprise AI in India.
"These models can be deployed on air-gapped environments, which means enterprise-based artificial intelligence, which thus far hasn't really taken off, would be possible to do a lot more of that in the country using models like this," said S. Krishnan.
By removing the need for cloud connectivity, these models open doors for highly secure applications in sectors such as:
- Cybersecurity: Protecting sensitive data from external threats.
- Scientific Research: Processing proprietary data in secure labs.
- Government Operations: Handling classified information without risk of leaks.
- Mathematics and Coding: Specialized tools for technical productivity.
A National Resource for All
The government is not keeping these tools limited to high-end corporations. To ensure a democratic distribution of technology, the models are accessible via AIKosh. This platform makes the AI models available to a wide spectrum of users, including:
- Students and Researchers
- Innovators and Startups
- Enterprises
- The general public
Secretary Krishnan concluded by stating that smaller, efficient models would drive productivity and efficiency across diverse applications, and the government remains committed to the ongoing refinement and development of these indigenous tools.