IndiaAI Mission Released Indigenous Foundation Models EKA-1 by Soket and Dhīr 2B & Dhīr 5B by ZenteiQ Under the IndiaAI Innovation Centre Pillar

1. At a Glance

• What happened: The IndiaAI Mission (MeitY) released three home-built foundation models: EKA-1 by Soket, and Dhīr 2B and Dhīr 5B by ZenteiQ. They were built under the IndiaAI Innovation Centre (Foundation Models) pillar [1].
• The policy model: The state provides subsidized public AI compute and pays for ancillary research expenses. Private Indian developers then build the models [1].
• Two kinds of model: EKA-1 is a large, enterprise-grade model covering all 22 scheduled languages [1]. Dhīr is lightweight, works on-device or on-premise, and supports 10 Indian languages plus English [1].
• UPSC relevance: GS-III (S&T, indigenisation of technology, cybersecurity), GS-II (government policies), and the wider debate on technological and strategic autonomy (sovereign AI) [1].

2. Why in the News

• 03 Nov 2026: IndiaAI Mission released EKA-1, Dhīr 2B and Dhīr 5B [1].
• The models were released in New Delhi by the MeitY Secretary, S. Krishnan, according to a PIB photo caption [2].
• The MeitY Secretary said indigenous models strengthen technological and strategic autonomy and improve productivity in government, industry, research and education [1].
• Earlier in 2026, models from Sarvam AI, BharatGen, Gnani and Soket were launched at the India AI Impact Summit 2026 [6].

3. Background & Evolution

• Why the mission exists: India depended on foreign foundation models. The mission's vision is "Making AI in India and Making AI Work for India" [1].
• Timeline:
• 7 Mar 2024: Union Cabinet approved the IndiaAI Mission with an outlay of ₹10,371.92 crore over 5 years [3].
• 30 Jan 2025: Call for proposals for foundation models; 506 applications received [3].
• Phase 1: Four startups selected: Sarvam AI, Soket AI, Gnani AI, Gan AI [3].
• AIKosh launched by MeitY as a repository of datasets, models and use cases, with an AI sandbox [5].
• 20 foundation-model proposals identified for support: 12 Large Multimodal Models (LMMs) and 8 Small Language Models (SLMs) [3].
• Feb 2026: 12 startups selected across the first two phases [7].
• Feb 2026: India AI Impact Summit; Sarvam, BharatGen, Gnani and Soket models launched [6].

• 03 Nov 2026: EKA-1 and Dhīr 2B/5B released [1].

• Related earlier work:

• BharatGen model (June 2025) supports 22 Indian languages with text, speech and image understanding [6].
• Sarvam received approval for India's Sovereign LLM in April 2025 [6].

4. Core Static Facts

5. Multi-Dimensional Analysis

• Scientific / Technological
• A sparse hybrid design (EKA-1) aims at efficiency without losing reasoning or agentic ability [1].
• Small, on-device models (Dhīr) make AI usable without cloud access, i.e. at the edge [1].

• Coverage of all 22 scheduled languages addresses the gap foreign models leave in Indic languages [1].

• Geopolitical / Strategic

• Indigenous models reduce dependence on foreign models and build strategic autonomy [1].

• Offline, on-premise deployment suits defence and sensitive government use, where data cannot leave the premises [1].

• Economic

• Subsidized compute lowers the entry barrier for domestic startups [1].
• EKA-1's BFSI focus targets a sector central to digital finance [1].

• Expected productivity gains across government, industry, research and education [1].

• Social / Inclusion

• Covering scheduled languages widens access for non-English speakers [1].
• Phone-capable models reach low-connectivity areas [1].

• Open release via AIKosh opens access to students, researchers and innovators [1].

• Ethical / Governance

• On-device processing improves privacy and data security [1].
• The Safe & Trusted AI pillar and the India AI Governance Guidelines (Feb 2026) frame responsible deployment [1][8].
• Gap: There is no confirmed public benchmark data yet for these specific models. Independent evaluation remains an open issue (analytical point).

6. Recent Developments (last 12–18 months)

• Feb 2026: India AI Impact Summit; Sarvam, BharatGen, Gnani and Soket models launched [6].
• Feb 2026: 12 startups in foundation-model phases 1–2 [7]; AIKosh at 7,541 datasets and 273 models [7].
• Feb 2026: India AI Governance Guidelines released [8].
• Mar 2026: PIB reported that the sovereign models perform strongly on Indic-language benchmarks, and Sarvam models are on AIKosh [6].
• 03 Nov 2026: EKA-1 and Dhīr 2B/5B released [1][2].

7. Prelims Hooks

• IndiaAI Mission approved by the Union Cabinet on 7 March 2024 [1].
• Mission outlay: ₹10,371.92 crore over five years [3].
• Nodal ministry: MeitY, not DST or NITI Aayog [1].
• IndiaAI Mission has seven pillars [1].
• EKA-1 was developed by Soket; Dhīr 2B and 5B by ZenteiQ [1].
• EKA-1 has 26.3 billion parameters [1].
• EKA-1 was trained on ~7.5 trillion tokens [1].
• EKA-1 covers all 22 scheduled languages (Eighth Schedule) plus English [1].
• Dhīr supports 10 Indian languages plus English [1].
• Dhīr runs on GPUs, TPUs and mobile phones and is offline-capable [1].
• EKA-1 is described as an open, sparse hybrid model for math, code, reasoning and agentic use [1].
• AIKosh is the IndiaAI Datasets Platform pillar [1].
• The foundation-model call for proposals (30 Jan 2025) drew 506 applications [3].
• Phase-1 startups: Sarvam AI, Soket AI, Gnani AI, Gan AI [3].

8. An Indian Model That Still Runs on Foreign Chips

• The model is Indian, but the hardware it is trained on is not
• A GPU (graphics processing unit) is the chip used to train and run AI models.
• The government told Parliament that GPUs are highly advanced equipment and are made mainly in one country [9].

• So India owns EKA-1 and Dhīr, but every GPU-hour spent training them was bought from abroad. If that supply is cut off or restricted, India could not train the next version at home.

• Most of the compute is rented, not owned

• The mission uses 15 empanelled compute service providers (private companies that rent out GPU time) [9].
• The state's own capacity is much smaller. One purchase order covers a system of about 1.1 EFLOPS (a measure of computing speed) at the NIC Data Centre in Delhi [9]. A secure national cluster of 3,000 GPUs is being built for sovereign and strategic uses [9].

• This matters for Dhīr. It is aimed at defence and sensitive government work, but that work depends on the small, state-controlled part of the compute, not the large rented part.

• The home-grown chip fix is still at the design stage

• Under the National Supercomputing Mission (NSM), work has started on Indian processors and GPGPUs (general-purpose GPUs) [9].
• These are based on RISC-V, an open-source instruction set (a free, shared rulebook for designing chips) [9]. No licence is needed, so a foreign company cannot withhold it.

• Projects that have just started will not replace imported GPUs for years. For now, "strategic autonomy" covers the model and the data, not the chip.

• Who should act: MeitY should plan the AI layer and the India Semiconductor Mission together. The OECD's 2023 compute blueprint treats resilience (security, sovereignty, sustainability of supply) as its own test for any national compute plan, separate from raw capacity [13].

9. Who Checks Whether These Models Are Any Good?

• Right now, the performance claim comes from the people who paid for the models
• PIB has said the sovereign models perform strongly on Indic-language benchmarks [6]. A benchmark is a standard test used to score and compare AI models.

• That claim comes from the government, which also funded the models. No independent Indian body has published a standard test that these models were scored on.

• The government's own guidelines say the testing yardstick still has to be built

• The India AI Governance Guidelines ask for the IndiaAI Safety Institute to get enough resources to draft standards, evaluation metrics, testing methods and benchmarks [8].
• The guidelines also point to the need for standardised "evaluation datasets" (fixed test questions that every model must answer) [8].

• So the models were released before the test meant to judge them exists.

• The sectors these models target need the most checking

• EKA-1 is pitched at BFSI and cybersecurity [1]. In a bank or insurer, a confident wrong answer leads to a wrong loan or claim decision.

• Dhīr is pitched at defence and offline use [1]. A model that runs on-premise, with no cloud link, does not get quick central fixes once it is installed. Each copy has to be updated separately.

• Who should act: The IndiaAI Safety Institute should publish an open Indic benchmark covering accuracy, safety and bias across all 22 scheduled languages, as the Guidelines recommend [8]. All 12 funded startups [7] should then report results on it before their models are used in government.

10. Smaller and Later Than First Promised: Is That a Problem?

• The targets moved
• A 2025 PIB release headline said a national large language model was slated for launch by end-2025 [10]. The first sovereign models came out at the February 2026 summit [6]. EKA-1 came out in November 2026 [1].

• Earlier PIB material described Soket aiming for a model of about 120 billion parameters. EKA-1 has 26.3 billion [1] (see Section 10 of the note).

• The strongest criticism

• The world's leading models are far larger, and they keep improving.

• A critic would say a 26-billion-parameter model will be out of date quickly. On this view, India could save public money by taking open foreign models and fine-tuning them (training them further) on Indian data.

• What the critic gets right

• With subsidised, rented compute [9], India is not competing to build the world's biggest model. The note should not suggest otherwise.

• The missed deadline [10] is a fair point. Ministries should not announce dates they cannot meet.

• Why the smaller-model choice still makes sense

• A fine-tuned foreign model is still a foreign model. Its owner decides the licence, the training data and the next update. Owning the weights (the learned settings that make up a model) removes that dependence. That is the autonomy argument the Secretary made [1].
• Small models fit Indian conditions. Dhīr runs on mobile phones and works offline [1]. A frontier-sized model cannot do that, so it would not reach low-connectivity districts.

• EKA-1's sparse design [1] means only part of the model works on each task. That keeps running costs low, which matters to a bank or a government office paying for every query.

• The honest verdict: India has bet on models that are useful and affordable, not on being the biggest. That bet is reasonable. Whether it works can only be judged once the independent testing described above exists.

11. GPU Counts Show What Was Sanctioned, Not What Was Used

• The headline figures count GPUs available, not GPUs used
• Official figures for common compute have moved from 34,000 GPUs [4] to 38,000 GPUs at about ₹65 per hour [10], and then to 45,000+ GPUs by June 2026 [11].

• These figures count GPUs onboarded (made available). They do not show how many hours those GPUs actually ran.

• The usage figure is a promise, not a measurement

• The government told Lok Sabha (5 August 2026) that 237 projects have been approved for subsidised compute, with 93.18 lakh GPU hours sanctioned [9].

• "Sanctioned" means allotted. If a startup does not use its hours, the subsidy is paid out without a model at the end. The public cannot see whether this is happening.

• What Canada did differently

• Canada's AI strategy sets aside CAD 1 billion to build sovereign compute. It adds a separate CAD 300 million Access Fund so that small firms and researchers can afford to use it [12].

• Building capacity and paying for its use are two separate budget lines there. So each can be tracked on its own.

• Who should act: MeitY should publish GPU hours actually used, per funded model and per provider, next to the sanctioned figure. The OECD blueprint asks countries to measure compute by capacity (what exists and is used), effectiveness (who can access it and what it produces) and resilience [13]. India's reporting covers only the first, and only half of that.

12. Anchors for Answers

• Data: 93.18 lakh GPU hours sanctioned across 237 projects under subsidised compute (Lok Sabha, 05.08.2026). This is hours allotted, not hours used [9]
• Data: Common compute grew from 34,000 GPUs [4] to 45,000+ GPUs by June 2026 [11]. The state is also building its own 3,000-GPU secure cluster for strategic use [9]
• Report/Committee: India AI Governance Guidelines (2026). They ask that the IndiaAI Safety Institute be resourced to build benchmarks and evaluation datasets [8]
• Report/Committee: OECD, A Blueprint for Building National Compute Capacity for AI (2023). It tests compute by capacity, effectiveness and resilience [13]. Also OECD, Securing AI Compute: Government Policy Approaches (Oct 2026) [12]
• Law/Case: Eighth Schedule of the Constitution (22 languages), the basis of EKA-1's language coverage [1]. DPDP Act, 2023, relevant to Dhīr's on-device data handling
• Comparison: Canada set aside CAD 1 billion for sovereign compute, plus a separate CAD 300 million Access Fund for SMEs and researchers. Building compute and paying for its use are funded separately [12]
• Scheme: National Supercomputing Mission, where RISC-V-based Indian processors and GPGPUs are being designed. This is the long-term answer to imported GPUs [9]

13. Mains Relevance

• GS-III:
• Science & Technology: developments and their applications; indigenisation of technology.
• Awareness in IT and computers.

• Challenges to internal security through communication networks; cybersecurity.

• GS-II: Government policies and interventions; e-governance.

• GS-IV (peripheral): Ethics of AI, privacy and accountability.
• Possible question stems: 1. "Public compute infrastructure is the backbone of India's sovereign AI ambitions." Examine this in light of the IndiaAI Mission's foundation-model pillar. 2. Discuss how small, on-device language models can make AI more inclusive and secure in India. What challenges remain? 3. "Technological autonomy in AI is a strategic necessity, not a choice." Critically analyse with reference to recent indigenous LLM initiatives.

14. Related Topics to Study Next

• IndiaAI Compute Portal / GPU subsidy: the input that made these models possible.
• AIKosh: where the models and datasets are distributed.
• BharatGen and Bhashini: earlier government-backed multilingual AI and language-technology work.
• India AI Impact Summit 2026: the global positioning of India's AI agenda.
• India AI Governance Guidelines (2026) and the DPDP Act, 2023: regulation, privacy and safe AI.
• India Semiconductor Mission: the hardware layer of sovereign AI.
• Eighth Schedule (22 languages): the basis of the language-coverage claims.

15. NCERT Concepts to Revise

• Strategic autonomy (Polity): The Secretary framed indigenous models as building technological and strategic autonomy; this is the article's central policy argument.
• Eighth Schedule (Polity): EKA-1's coverage of all 22 scheduled languages rests on the Eighth Schedule's list of constitutionally recognised languages.
• India Semiconductor Mission (Geography): Models still run on imported GPUs; this mission is India's hardware-side answer to that dependence.
• Digital Personal Data Protection Act 2023 (Polity): Dhīr keeps data on the device, so the law governing consent and handling of personal data is relevant.
• Subsidy (Economy): The state subsidises compute for private developers; understanding subsidies helps judge whether sanctioned GPU hours deliver real models.
• Self-reliance (Economy): Building home-grown AI models to cut dependence on foreign ones applies the self-reliance idea to technology.

16. Common Errors / Trap Areas

• Parameters vs tokens: 26.3 billion is the parameter count; ~7.5 trillion is the training tokens [1].
• Which model covers what: EKA-1 covers 22 languages; Dhīr covers 10 [1].
• Soket's earlier 120B target: Earlier PIB material said Soket aimed for a ~120-billion-parameter model. EKA-1 is 26.3B, so don't conflate the two (based on search-result summaries).
• Spelling: Some PIB text writes "Socket". The firm is Soket (Soket AI) [1][6].
• Ministry and date: The ministry is MeitY, and the Cabinet approval was 7 March 2024, not the 2023 announcement or the Feb 2026 summit [1][3].