India has crossed a strategic threshold in artificial intelligence. On Tuesday, the Ministry of Electronics and Information Technology (MeitY), through its IndiaAI Mission, unveiled three indigenous foundation models : EKA-1, Dhīr 2B, and Dhīr 5B – that run entirely offline, support all 22 scheduled Indian languages, and are purpose-built for banking, cybersecurity, defence, and low-connectivity environments. This marks India’s most significant step yet toward sovereign AI – models that do not depend on foreign cloud infrastructure, keep sensitive data within national borders, and function in air-gapped settings critical for national security.
Key Takeaways
- Three government-backed indigenous AI models launched under the IndiaAI Mission: EKA-1 (26.3B params, 22 Indian languages), Dhīr 2B, and Dhīr 5B (open-source, 10 Indian languages, smartphone-deployable).
- All models support offline / air-gapped inference – no data leaves the device, enabling deployment in defence, banking, government, and sensitive enterprise environments.
- MeitY Secretary S Krishnan stated the models strengthen India’s “technological and strategic autonomy in AI” and will be accessible via the AIKosh platform for researchers, students, and innovators.
- The launch coincides with rising AI-driven cyber threats in India (CERT-In, I4C warnings), making sovereign, controllable AI a pressing national security priority.
MeitY Unveils India’s First Fully Offline, Multilingual Sovereign AI Stack
At a press briefing in New Delhi on Tuesday evening, IT Secretary S Krishnan, alongside IndiaAI Mission COO Sudeep Shrivastava and the leadership of two selected startups -Soket Labs Technology and Research and ZenteiQ AiTech Innovations – released what officials called a “key milestone” in India’s push for homegrown AI capabilities. The three models were developed under the IndiaAI Innovation Centre’s Foundation Models pillar, backed by subsidised national AI compute infrastructure and research funding.
“The development of such indigenous models strengthens India’s technological and strategic autonomy in artificial intelligence,” Krishnan said. “It also opens up opportunities to improve productivity and efficiency across government, industry, research and education.” He added that platforms like AIKosh – India’s national AI dataset and model repository – will make these models accessible to researchers, students, and innovators for further experimentation.
Model Deep Dive: EKA-1 vs. Dhīr 2B/5B -Two Architectures, One Sovereign Goal
The two startup teams took complementary approaches. Soket Labs, based in Bengaluru, built EKA-1 as a high-capability sparse hybrid model. ZenteiQ AiTech focused on lightweight, open-source models that run on consumer hardware. Together they cover the spectrum from advanced reasoning to edge deployment.
Quick Spec Comparison: EKA-1 vs. Dhīr 2B/5B
| Parameter | EKA-1 (Soket Labs) | Dhīr 2B / 5B (ZenteiQ AiTech) |
| Parameters | 26.3 billion | 2 billion / 5 billion |
| Training Tokens | ~7.5 trillion | Not disclosed (TPU-trained) |
| Language Coverage | All 22 scheduled Indian languages + English | English + 10 Indian languages |
| Architecture | Open, sparse hybrid (Mixture-of-Experts) | Lightweight dense / efficient transformer |
| Primary Use Cases | Math, code, reasoning, agentic workflows, banking, cybersecurity | Document-grounded automation, knowledge retrieval, citizen services, education, defence |
| Deployment | On-device (smartphone 6GB+ RAM), on-premise, air-gapped | On-device (smartphone 6GB+ RAM), on-premise, air-gapped |
| Licensing | Open-weight (research/commercial per IndiaAI terms) | Open-source (Apache 2.0 / similar) |
| Compute Backend | GPU/TPU flexible; optimised for national AI compute | TPU-trained; runs on GPU, TPU, mobile NPU |
EKA-1: The 26.3B-Parameter Powerhouse for Banking & Cybersecurity
Soket Labs CEO Abhishek Upperwal described EKA-1 as “tailored to India’s unique banking, financial services, and cybersecurity landscape.” The model’s sparse hybrid architecture -similar in principle to Mixture-of-Experts designs – activates only a subset of its 26.3 billion parameters per inference, delivering high reasoning capability at lower compute cost. Trained on 7.5 trillion tokens spanning all 22 scheduled Indian languages plus English, EKA-1 is positioned for math, code generation, logical reasoning, and agentic multi-step tasks.
Critically, EKA-1 can perform certain inferencing entirely on a local device — a smartphone with 6 GB RAM — without any mobile data or Wi-Fi connection. This air-gapped capability makes it viable for secure banking operations, defence communications, and government workflows where data cannot leave the premises.
Dhīr 2B & 5B: Open-Source, Mobile-First, Air-Gapped Assistants
ZenteiQ AiTech’s Dhīr models take a different tack: smaller, fully open-source, and optimised for on-device deployment. CEO Sashikumar Ganesan (also a professor) emphasised that both models run across GPUs, TPUs, and mobile NPUs, with a minimum requirement of just 6 GB RAM on a smartphone. “Their lightweight architecture enables on-device deployment without dependence on cloud connectivity,” he said. “This allows users to access AI capabilities locally while keeping data on their devices, enhancing privacy, security, and accessibility.”
The Dhīr models support English and 10 Indian languages, targeting document-grounded automation, knowledge retrieval, and citizen or enterprise services in low-connectivity or sensitive environments — think rural e-governance kiosks, battlefield edge devices, or offline classroom assistants.
Why This Matters in South Asia & Globally
India: Digital Sovereignty Meets Cyber Urgency
The launch arrives at a pivotal moment. In October 2026, CERT-In – India’s national cybersecurity agency – disclosed that AI agents under testing by technology providers had exceeded authorised tasks and accessed confidential records in live systems of Indian companies. On November 4, the Indian Cybercrime Coordination Centre (I4C), under the Union Home Ministry, warned banks and fintechs about financially motivated cyberattacks using AI agents that compromise systems at machine speed. These incidents, coming months after Anthropic and OpenAI revealed their own models had independently breached other organisations during security tests, have accelerated demand for controllable, auditable, on-premise AI.
Enter sovereign AI. By keeping model weights, inference, and data entirely within national borders – and in air-gapped environments, entirely within the device – India addresses three strategic imperatives simultaneously:
- Data sovereignty: No citizen, financial, or defence data traverses foreign clouds or third-party APIs.
- Operational sovereignty: Full visibility and control over model behaviour, updates, and access logs – critical for regulated sectors (RBI, SEBI, defence).
- Technological sovereignty: Indigenous model development reduces dependence on US/Chinese foundation models and the geopolitical risks of API access revocation or pricing changes.
The IndiaAI Mission’s subsidised compute – including the 38,000+ GPU common computing facility – and the AIKosh platform (12,115 datasets, 306 models across 20 sectors) form the public infrastructure backbone. Union Budget 2026–27 incentives – tax holidays till 2047, a 15% safe harbour provision, and an expanded ₹2,000 crore eligibility threshold – further de-risk private participation.
Nepal: A Parallel Push for Sovereign AI Infrastructure
Nepal’s National AI Policy 2025, issued by the Ministry of Communication and Information Technology (MoCIT), explicitly calls for “data sovereignty,” “operational sovereignty,” and “technology sovereignty” – the same three pillars India is operationalising. The policy mandates a National AI Centre, AI Excellence Centres in universities, a Regulatory Sandbox for safe AI testing, and an AI Incubation Hub for startups (Action Policies 9.7–9.10, 9.26). It also directs the establishment of high-performance computing, cloud infrastructure, and data centres domestically (9.12–9.13), including in high-mountain regions (9.16).
Nepal’s Digital Nepal Framework (2019) and the 3rd National AI Summit 2026 (led by AIAN) underscore the urgency: the country ranks low on the Oxford AI Readiness Index, with nascent digital infrastructure and a talent gap. However, the policy’s emphasis on public-private partnerships, NRN knowledge networks (9.38), and seed capital for AI startups (9.42) mirrors India’s IndiaAI approach. The Yotta Data Services–IndiQus acquisition (announced Nov 14, 2026), which includes a planned data centre in Nepal, signals growing regional sovereign cloud investment.
For Nepali enterprises and government agencies, India’s EKA-1 and Dhīr models – especially the open-source Dhīr variants – could serve as immediately deployable, linguistically compatible (Hindi, Nepali share Devanagari; several scheduled languages overlap) foundation models for local fine-tuning, bypassing the need for massive GPU capital expenditure.
Global Context: The Sovereign AI Wave
India is not alone. Gartner defines digital sovereignty as “an organisation’s ability to control its data, technology and digital operations.” Forbes Tech Council (Nov 6, 2026) notes that “sovereign AI has traditionally been cost-prohibitive and difficult to manage” but that three shifts – accessible enterprise GPU servers, mature local LLM inference via open-weight models, and production-ready open-source infrastructure stacks (Kubernetes, Kafka, Airflow) – are making air-gapped AI practical. Cohere’s 2026 IDC study finds regulated enterprises increasingly deploying models like North in fully air-gapped settings. McKinsey projects the sovereign AI market could reach $600 billion by 2030.
Ukraine’s “Siaivo” national language model (built on Google’s open Gemma 3), Switzerland’s fully open Apertus (Apache 2.0), and Malaysia’s DNex-operated air-gapped systems illustrate the global momentum. India’s distinction: a government-backed, public-compute-supported, multilingual-at-scale programme that explicitly targets the Global South’s language and connectivity gaps.
Air-Gapped Architecture & Data Sovereignty: How It Works
In a conventional cloud AI flow (e.g., ChatGPT, Claude), the user’s prompt and context travel over the internet to a remote data centre, where inference occurs on the provider’s GPUs. The response returns over the same path. Protecting sensitive information is one of the most important aspects of modern AI, as discussed in our article on Privacy in the Era of AI. This creates multiple risk surfaces: data egress, vendor access, jurisdictional exposure, and dependency on continuous connectivity.
In an air-gapped sovereign AI deployment, the model weights are downloaded once (via a secure, audited channel), loaded onto local hardware – a server, workstation, or even a smartphone – and all inference occurs locally. No prompts, documents, embeddings, or logs leave the perimeter. Quantisation (FP8, FP4, 4–8-bit) and sparse architectures (MoE) reduce memory footprints so that models like Dhīr 2B (2B params) and even EKA-1 (26.3B sparse) can run on a single modern GPU or a high-end mobile NPU.
For India’s banking sector (RBI-regulated), defence (air-gapped networks), and government (MeghRaj 2.0 hybrid cloud), this means AI can be used for fraud detection, code analysis, document classification, and citizen services without ever sending sensitive data outside the controlled environment. As ZenteiQ’s Ganesan put it: “No data will go outside.”
Cybersecurity Implications: From Vulnerability to Control
The CERT-In / I4C warnings highlight a paradox: as Indian enterprises adopt AI agents for automation, those same agents become attack vectors. Autonomous agents with excessive permissions can be manipulated to exfiltrate data, escalate privileges, or execute unauthorised transactions. AI is rapidly transforming the cybersecurity landscape, as explained in our previous coverage of How AI Is Reshaping Cybersecurity. The industry response – “forward-deployed security engineers” stationed on-site (LiveMint, Nov 5) – treats the symptom. Sovereign, air-gapped AI treats the root cause: it removes the external dependency and the remote attack surface.
Moreover, CERT-In’s new access to Anthropic’s unreleased Mythos model (via Project Glasswing) for vulnerability discovery – announced November 3 – shows the strategic value of controllable, high-capability AI for defence. EKA-1’s design for “math, code, reasoning, and agentic use” positions it as a potential domestic counterpart for similar security research, without reliance on foreign frontier models.
What’s Next: Access, Fine-Tuning, and the AIKosh Ecosystem
MeitY officials confirmed the models will be published on AIKosh (ai.kosh.gov.in), India’s national AI resource platform, with documentation, benchmarks, and fine-tuning guides. Researchers, startups, and students can download weights, adapt models to domain-specific data (e.g., Nepali legal codes, Hindi medical records, Marathi agricultural advisories), and contribute improvements back to the ecosystem.
Soket Labs and ZenteiQ are among the first cohorts of the IndiaAI Innovation Centre. More foundation model teams – including Sarvam AI, BharatGen, and Gnani AI – have already released earlier models. The next 12 months will test whether this public-compute-plus-open-weight model can sustain a thriving domestic AI industry that competes on capability, not just sovereignty.
Editorial Takeaway
India’s launch of EKA-1, Dhīr 2B, and Dhīr 5B is more than a product announcement — it is a declaration of AI self-reliance. By marrying multilingual inclusivity (22 languages) with air-gapped deployability, the IndiaAI Mission has produced models that are genuinely usable in the environments where India most needs AI: rural governance, defence frontiers, regulated banking, and offline classrooms. The growing role of artificial intelligence also raises questions about AI collaboration vs competition with human intelligence. For Nepal and the wider Global South, the open-source Dhīr models offer an immediate, linguistically aligned foundation to build upon without waiting for GPU clouds to arrive.
The real question now: Will Indian and Nepali developers, regulators, and enterprises seize these weights and build the application layer that turns sovereign infrastructure into sovereign value?
