{
  "id": 18008409,
  "title": "Two government-backed startups release sovereign AI models that can run in airgapped environment",
  "url": "https://urgent.news/2026/11/03/two-government-backed-startups-release-sovereign-ai-models-that-can",
  "topic": "ai",
  "section": "AI",
  "published": "2026-11-03T19:04:48.000Z",
  "source": {
    "name": "Hindustan Times",
    "slug": "hindustan-times",
    "url": "https://www.hindustantimes.com/india-news/two-startups-release-sovereign-ai-models-that-can-run-in-airgapped-environment-101793732689316.html"
  },
  "original_language": "en",
  "account": "Government-backed startups Soket AI and ZenteiQ have unveiled sovereign AI models that can operate in air-gapped environments, which are completely isolated from the internet and external networks. The models were introduced at a press conference in New Delhi, attended by IT Ministry secretary S Krishnan and COO of India AI Mission Sudeep Shrivastava.\n\nSoket AI introduced EKA-1, a 26.3 billion parameter model designed for mathematics and coding. This model was trained on approximately 7.5 trillion tokens, encompassing all 22 Indian languages and English. ZenteiQ, on the other hand, released two open-source models, Dhīr 2B and Dhīr 5B, which are multilingual assistants supporting English and ten Indian languages. These models are intended for secure, on-premise, and offline use in critical sectors.\n\nZenteiQ's Sashikumar Ganesan explained that air-gapped systems differ from popular chatbots like ChatGPT. In a service like ChatGPT, the \"inferencing\" - the process where the model generates an answer - takes place on a remote server. However, in an air-gapped setup, the inferencing occurs locally, meaning no data is transmitted outside the system. Ganesan also highlighted that the models can run on a mobile phone with a minimum of 6 GB of memory, and that they can be deployed in various sectors, including defence and cybersecurity.\n\nIT Ministry secretary S Krishnan commented that smaller models have numerous applications, including cybersecurity, and can be effectively utilized in air-gapped environments. He noted that Soket AI's model is optimized for mathematics and coding, while ZenteiQ's model is geared towards science and research. Ganesan explained that ZenteiQ trained their models on Tensor Processing Units (TPUs) due to the limited availability of Graphics Processing Units (GPUs). He clarified that GPUs are not necessarily superior to TPUs and that acquiring GPUs would have resulted in a delay of three to four months.",
  "summary": "India AI Mission-backed startups Soket AI and ZenteiQ unveiled sovereign, offline models for secure use in coding, maths, science and research.",
  "key_points": [
    "Government-backed startups Soket AI and ZenteiQ release air-gapped AI models.",
    "EKA-1, a 26.3 billion parameter model, designed for mathematics and coding.",
    "Dhīr 2B and Dhīr 5B models support English and ten Indian languages for secure use."
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}