# Two Government-Backed Startups Release Sovereign AI Models Under IndiaAI Mission for Air-Gapped Environments Source: TechNewsList (https://technewslist.com) Canonical URL: https://technewslist.com/en/article/two-startups-release-sovereign-ai-models-under-indiaai-missio-2026-11-04-morning Section: AI (https://technewslist.com/en/ai) Author: TechNewsList Language: en Published: 2026-11-04T06:09:36.862+00:00 Updated: 2026-11-04T06:09:36.865339+00:00 > India's Ministry of Electronics and IT launches the first sovereign foundation models, EKA-1 and Dhir, engineered for offline execution in air-gapped defense, banking, and citizen administration environments. ## TL;DR - MeitY Secretary S. Krishnan officially released India's first wave of sovereign foundation models under the government-backed IndiaAI Mission. - Soket AI Labs introduced EKA-1, a 26.3-billion-parameter sparse hybrid model trained on 7.5 trillion tokens spanning all 22 scheduled Indian languages. - ZenteiQ AiTech Innovations launched Dhir 2B and 5B, lightweight assistant models optimized for on-device inferencing on edge hardware. - Both model families are engineered to operate in air-gapped, offline environments to preserve national security and data sovereignty. ## Key points - EKA-1 is tailored specifically for complex mathematical reasoning, software development, and sovereign automated workflows in banking and cybersecurity. - Dhir models can execute locally on commercial smartphones with as little as 6GB of RAM, enabling AI access in low-connectivity rural regions. - The initiative reduces strategic reliance on Western and Chinese cloud APIs for critical government, military, and financial services. - All training datasets were curated to prevent unauthorized cross-border telemetry and comply with strict national data residency mandates. - MeitY plans to distribute model weights directly to state government agencies, public universities, and domestic enterprise partners. ## What happened In a landmark milestone for India's technological autonomy, the Ministry of Electronics and Information Technology (MeitY) officially rolled out the nation's first suite of indigenous sovereign foundation artificial intelligence models. Unveiled during an executive ceremony in New Delhi led by MeitY Secretary S. Krishnan, the release represents the initial tangible output of the government-funded IndiaAI Mission, which was established to develop sovereign computing infrastructure and reduce strategic dependence on foreign cloud platforms. The launch encompasses foundation models engineered by two government-backed domestic artificial intelligence startups: Soket AI Labs and ZenteiQ AiTech Innovations. Soket AI Labs introduced EKA-1, an advanced 26.3-billion-parameter open, sparse hybrid language model trained across an expansive dataset of 7.5 trillion tokens. EKA-1 is designed to support high-order mathematical reasoning, software development, agentic task automation, and high-security transactional processing across banking and cybersecurity. Simultaneously, ZenteiQ AiTech Innovations debuted its Dhir model family, offering open-weight multilingual assistants in compact 2-billion and 5-billion parameter configurations. Unlike massive hyperscale models that require multi-node GPU clusters, Dhir is engineered to run locally on resource-constrained consumer edge devices, including Android smartphones equipped with at least 6 gigabytes of memory, ensuring widespread accessibility across rural administrative districts. ## Why it matters Global geopolitics and cybersecurity mandates have elevated digital sovereignty from an abstract policy concept into an urgent national defense imperative. Most contemporary frontier models operate within proprietary cloud environments located in North America or East Asia. Routing sensitive government deliberations, intelligence summaries, judicial records, and citizen personal data through foreign API endpoints introduces unacceptable surveillance and supply chain vulnerabilities for sovereign nations. By ensuring that both EKA-1 and the Dhir model series can execute entirely within air-gapped physical environments—systems physically isolated from external local networks and the public internet—India provides its defense forces, state intelligence bureaus, and central banking infrastructure with modern natural language capabilities that guarantee complete data residency. ![Software development environment showing programming scripts and multilingual tokenization pipelines.](https://technewslist.com/storage/v1/object/public/media/api/1793772572916-6pfd48-two-startups-release-sovereign-ai-models-under-indiaai-missio-2026-11-04-morning-inside-1-d1c0d82a74.webp) Furthermore, the models bridge a massive linguistic divide. Mainstream global foundational models are disproportionately trained on English and high-resource European languages, frequently exhibiting degraded reasoning, cultural hallucination, or poor grammatical syntax when processing Indian regional dialects. EKA-1 and Dhir natively comprehend and generate text across all 22 scheduled Indian languages, democratizing automated administrative services for hundreds of millions of citizens. ## Technical details The architectural design of EKA-1 reflects modern innovations in parameter-efficient compute. Rather than utilizing a dense transformer topology that activates every parameter during inference, Soket AI Labs implemented a sparse hybrid architecture incorporating mixture-of-experts routing layers. This allows the 26.3-billion-parameter model to achieve inference speeds and memory bandwidth profiles comparable to a dense 7-billion-parameter network while preserving the reasoning depth of a far larger model. The training corpus for EKA-1 was curated specifically to eliminate western cultural bias while prioritizing regional administrative terminology, legal jurisprudence, and technical documentation across languages including Hindi, Tamil, Telugu, Bengali, Marathi, and Gujarati. The tokenizer was trained with expanded regional vocabulary allocations, dramatically reducing token fragmentation and computational inference latency for non-Latin scripts. ![Modern integrated development environment used for training, debugging, and deploying edge artificial intelligence weights.](https://technewslist.com/storage/v1/object/public/media/api/1793772573940-ftj8xr-two-startups-release-sovereign-ai-models-under-indiaai-missio-2026-11-04-morning-inside-2-b41cd8d531.webp) ZenteiQ's Dhir models utilize aggressive 4-bit and 8-bit quantization-aware training techniques, allowing full model weights to be mapped into edge device memory without degrading semantic coherence. The models support cross-architecture execution, running efficiently across x86 CPUs, ARM Cortex processors, Qualcomm mobile silicon, and enterprise GPU accelerators, providing versatile deployment options for state ministries. ## Market / industry impact The IndiaAI Mission's model rollout will accelerate domestic enterprise adoption of generative artificial intelligence across heavily regulated sectors. Indian commercial banks, non-banking financial companies, and healthcare providers have historically moved cautiously regarding generative AI deployment due to strict Reserve Bank of India data localization mandates. Having access to certified, domestically trained, air-gapped models eliminates these regulatory hurdles. The launch also establishes a competitive benchmark for indigenous startup ecosystems across the Global South. Rather than merely licensing proprietary model APIs or fine-tuning existing Western open-source weights, India's public-private venture strategy demonstrates how emerging economies can pool sovereign computational compute and curated public datasets to build foundational intellectual property. Domestic systems integrators and enterprise software vendors—including Tata Consultancy Services, Infosys, and Wipro—are already preparing sovereign implementation frameworks to deploy EKA-1 within municipal government automation projects, tax compliance portals, and state utility management systems. ## What to watch next Following the initial unveiling in New Delhi, MeitY will initiate pilot deployment programs across select central ministries, public sector defense manufacturing plants, and nationalized banking institutions. These trials will evaluate real-world latency, domain-specific accuracy, and air-gapped operational reliability under high transaction volumes. In parallel, the IndiaAI Mission plans to release open-access model weights and technical benchmark whitepapers to accredited academic institutions and independent software developers across India. This release is expected to trigger a wave of fine-tuned domain models specialized for local agricultural advisory, telemedicine, and multilingual educational tutoring. Finally, observers will track upcoming hardware procurement tenders under the IndiaAI Mission's subsidized compute pillar. As the government prepares to deploy thousands of additional high-performance GPUs across national datacenter facilities, sovereign infrastructure capacity will expand to support the next training iterations of EKA-2 and advanced multimodal vision models. ## Sources - [Hindustan Times](https://www.hindustantimes.com/india-news/two-startups-release-sovereign-ai-models-that-can-run-in-airgapped-environment-101793732689316.html) - [Ministry of Electronics and Information Technology](https://en.wikipedia.org/wiki/Ministry_of_Electronics_and_Information_Technology) - [Artificial Intelligence in India](https://en.wikipedia.org/wiki/Artificial_intelligence_in_India) Mentions: IndiaAI Mission, Soket AI Labs, ZenteiQ AiTech Innovations, Ministry of Electronics and Information Technology, S. Krishnan, New Delhi ## Sources - [Hindustan Times](https://www.hindustantimes.com/india-news/two-startups-release-sovereign-ai-models-that-can-run-in-airgapped-environment-101793732689316.html) - [Ministry of Electronics and Information Technology](https://en.wikipedia.org/wiki/Ministry_of_Electronics_and_Information_Technology) - [Artificial Intelligence in India](https://en.wikipedia.org/wiki/Artificial_intelligence_in_India)