India is shifting from a consumer of global AI technologies to a core architect. The IndiaAI Mission recently drove this change by unveiling three homegrown foundational AI models.
Soket Labs developed Eka-1, while ZenteiQ built Dhir 2B and Dhir 5B. This public release marks a defining moment for India’s digital sovereignty.
It shifts focus toward building culturally aware machine learning systems. These systems natively understand India’s linguistic diversity and unique institutional needs.
At the launch event, IT Secretary S. Krishnan spoke about this technological milestone. He noted these foundational models establish a robust, strategic AI ecosystem within India. The government deliberately targets practical, secure deployments instead of the generic generative AI race.
These indigenous models offer extensive potential across diverse sectors, poised to significantly enhance productivity, fortify cybersecurity, and streamline operational efficiency across both public administration and private enterprises.
Engineering the Multilingual Eka-1 Framework
When evaluating the underlying architecture of Eka-1, the sheer computational scale and linguistic inclusivity immediately stand out. Built by Soket Labs as an open, sparse hybrid language model, Eka-1 is powered by an impressive 26.3 billion parameters and has been exhaustively trained on approximately 7.5 trillion tokens.
This massive foundational training allows the model to perform highly complex functions, spanning mathematics, coding, advanced logical reasoning, and specialized agentic workflows.
However, what truly differentiates Eka-1 from its heavily Western-centric counterparts is its deep linguistic integration. It demonstrates highly advanced capabilities across all 22 scheduled Indian languages alongside English.
According to a recent broadcast update on the initiative at Newson air, the introduction of these specific systems is designed to dramatically strengthen homegrown AI capabilities from the ground up.
By embedding regional languages natively into the model’s earliest training phases rather than relying on superficial translation layers, Eka-1 generates vastly more nuanced, culturally accurate outputs.
This localized approach is absolutely vital for banking, financial services, and civic applications looking to expand digital accessibility across India’s vast, multilingual demographic.
Deploying Dhir 2B and 5B for Secure Operations
Eka-1 focuses on massive linguistic scale, but the Dhir models prioritize secure, localized infrastructure. ZenteiQ engineered Dhir 2B and Dhir 5B for offline-capable, document-grounded automation and precision knowledge retrieval.
This tailored functionality makes them exceptionally valuable for defense, government, industrial, and educational services. Highly regulated fields demand strict data privacy and uninterrupted operational continuity. Consequently, these organizations cannot safely rely on continuous external cloud connectivity.
Mr. Krishnan emphasized their strategic importance during his launch speech. He noted organizations can successfully deploy the Dhir models in entirely air-gapped environments. They run on isolated hardware without internet access, which drastically reduces external cyber threats.
Dhir 2B and 5B operate flawlessly in low-connectivity or highly sensitive environments. They ensure classified or proprietary data never leaves an organization’s internal servers. Ultimately, they bridge the crucial gap between cutting-edge intelligent automation and rigorous national security protocols.
