India’s AI Ambitions Get a Secure Boost: Indigenous Sovereign Models Launched for Air-Gapped Use
In a significant stride toward achieving technological self-reliance, the Indian government’s India AI Mission has showcased the successful development of two sovereign artificial intelligence models. Startups Soket AI and ZenteiQ unveiled their cutting-edge innovations on Tuesday in New Delhi, introducing systems designed specifically for “air-gapped” environments—computing networks that are entirely disconnected from the internet and external digital interfaces.
The unveiling, attended by IT Ministry Secretary S. Krishnan and India AI Mission COO Sudeep Shrivastava, marks a pivotal moment for domestic AI development, particularly for critical sectors where data security and sovereignty are paramount.
Bridging the Gap with Sovereign Intelligence
The core advantage of these new models lies in their ability to perform “inferencing”—the process of generating responses—locally on hardware rather than relying on remote, cloud-based servers. Unlike global commercial chatbots such as ChatGPT, which necessitate a continuous internet connection to fetch data from distant data centers, the models developed by Soket AI and ZenteiQ operate in total isolation.
Soket AI introduced EKA-1, a powerful model boasting 26.3 billion parameters. Engineered primarily for complex tasks like advanced mathematics and computer programming, EKA-1 represents a robust linguistic achievement. It has been trained on a massive dataset of 7.5 trillion tokens, incorporating all 22 scheduled Indian languages alongside English, ensuring the model is inclusive of India’s linguistic diversity.
Simultaneously, ZenteiQ launched its Dhīr series, consisting of two open-source models, Dhīr 2B and Dhīr 5B. These assistants are tailored for multilingual support, covering English and 10 Indian regional languages. Designed for high-security environments, these models are optimized for on-premise deployment, meaning sensitive data never leaves the facility where it is processed.
Strategic Utility for National Security
The operational capacity of these models extends to hardware as modest as a smartphone with 6 GB of RAM, making them highly portable and versatile. During the briefing, Sashikumar Ganesan of ZenteiQ emphasized that the air-gapped nature of these models makes them ideal for sectors such as national defense, cybersecurity, and sensitive scientific research. By eliminating external connectivity, the risk of data exfiltration or external cyber-tampering is effectively neutralized.
IT Ministry Secretary S. Krishnan lauded the diverse capabilities of the startups, noting that while Soket’s EKA-1 is focused on rigorous computational and coding logic, ZenteiQ’s models are optimized for broader scientific and research applications.
Overcoming Infrastructure Hurdles
A notable aspect of the development process was the startups’ ingenuity in overcoming global hardware shortages. Facing a scarcity of high-demand Graphics Processing Units (GPUs), ZenteiQ turned to Tensor Processing Units (TPUs) to train their models. Ganesan noted that this shift not only bypassed a three-to-four-month wait for GPUs but also proved that alternative architectures are fully capable of handling sophisticated AI workloads without sacrificing performance.
Addressing concerns regarding safety and ethics, Soket founder Abhishek Upperwal confirmed that the EKA-1 model has been equipped with comprehensive safety guardrails. These built-in protections are designed to filter and prevent the generation of harmful content, including material related to self-harm and suicide, ensuring the models remain responsible tools for national progress. As these sovereign models move toward wider implementation, they represent a secure, indigenous foundation for India’s rapidly evolving AI landscape.
Disclaimer: This content is auto-generated for informational purposes only.
Source: Read Original News
