IndiaAI Mission EKA-1 and Dhir Models: What MeitY Released, Specs and Why It Matters, Explained

The IndiaAI Mission, under the Ministry of Electronics and Information Technology (MeitY), on 3 November 2026 released indigenous foundation models EKA-1 and Dhīr (Dhīr 2B and Dhīr 5B). PIB says EKA-1 was developed by Soket Labs Technology and Research Private Limited and the Dhīr models by ZenteiQ AiTech Innovations. The release was unveiled at a press event in New Delhi. Here is what the government put on record, what the companies add, and what is still unpublished.

The 60-second version

• When and where: 3 November 2026, New Delhi. PIB Delhi posted the release at 10:02 pm. The Economic Times Online last updated its story at 11:08 pm IST. Moneycontrol's story is timestamped 10:57 pm IST.
• Who was there: MeitY Secretary S. Krishnan, senior MeitY and IndiaAI officials, and leadership from Soket and ZenteiQ (PIB).
• EKA-1: 26.3 billion parameters, trained on about 7.5 trillion tokens. PIB calls it an open, sparse hybrid language model for all 22 scheduled Indian languages plus English, aimed at math, code, reasoning and agentic use.
• Dhīr 2B and Dhīr 5B: Open-source, English plus 10 Indian languages, pitched for secure on-premise and offline use in defence, government, industry and education (PIB).
• Company specs: ZenteiQ's site says Dhīr 2B is a dense model of about 2.03 billion parameters, and Dhīr 5B is a mixture-of-experts model of about 5.09 billion total parameters with about 1.79 billion active per token. Both list a 16,384-token context.
• Released count: Moneycontrol says this release takes models released to five out of 20 backed by IndiaAI. It names BharatGen's Param 2, Gnan.ai's Vachana and a Sarvam AI suite as already out.
• Access: Krishnan pointed to AIKosh for researchers, students and innovators. The note says the models can be made accessible. It does not attach a download link.

What MeitY released, and who said what

PIB describes the models as a milestone under the IndiaAI Mission and says Soket and ZenteiQ are among the teams selected under the IndiaAI Innovation Centre (Foundation Models) pillar to build indigenous foundation models and large language models. It says access to subsidized national AI compute, and support for ancillary research expenses, enabled the work.

Krishnan said: "The development of such indigenous models strengthens India's technological and strategic autonomy in Artificial Intelligence. It also opens up opportunities to improve productivity and efficiency across government, industry, research and education." He added that EKA-1 and Dhīr address distinct use cases, and that through platforms such as AIKosh the models "can be made accessible to researchers, students, innovators and the wider public, encouraging further experimentation and innovation."

IndiaAI chief operating officer Sudeep Shrivastava said the mission's support is meant to make artificial intelligence "more accessible and tailored to India's unique landscape." He called Dhīr and EKA-1 evidence of progress "ranging from lightweight solutions optimized for everyday devices to robust powerhouses engineered for complex enterprise applications."

The Economic Times, last updated at 11:08 pm IST on 3 November 2026, also reported the launch and the presence of Krishnan and senior ministry and IndiaAI officials. It quoted ZenteiQ CEO Professor Sashikumaar Ganesan: "We have launched two small models. We call them compact models...what we have launched is a two-billion and then five-billion parameter model..." PIB spells the same CEO's name as Prof. Sashikumar.

EKA-1: 26.3 billion parameters, 22 languages

PIB says Soket's EKA-1 has 26.3 billion parameters and was trained on approximately 7.5 trillion tokens. It "demonstrates advanced capabilities, spanning all 22 scheduled Indian languages alongside English." The release calls it an open, sparse hybrid language model designed for math, code, reasoning, and agentic use. It does not define "sparse" or "hybrid", and it does not publish benchmark scores, an active-parameter count, or a licence text.

Soket CEO Abhishek Upperwal said: "With the support of the IndiaAI Mission, we have developed an advanced AI model specifically tailored to India's unique banking, financial services and cybersecurity landscape. It is designed to address sector-specific requirements, understand India's evolving digital ecosystem, and deliver solutions aligned with the country's needs and priorities."

That is a company description of the intended market. PIB does not list bank or cybersecurity deployments.

Dhīr 2B and Dhīr 5B: offline, and the specs on ZenteiQ's site

PIB says ZenteiQ's Dhīr 2B and Dhīr 5B are open-source models that can be deployed as secure, on-premise, multilingual assistants supporting English and 10 Indian languages. It says they enable offline-capable, document-grounded automation, knowledge retrieval, and citizen or enterprise services across defence, government, industrial and educational sectors, "particularly in low-connectivity or sensitive environments."

Prof. Sashikumar said: "Models like Dhīr 2B and Dhīr 5B offer the flexibility to run across different computing platforms, including GPUs and TPUs. Their lightweight architecture also enables on-device deployment on mobile phones, without dependence on cloud connectivity. This allows users to access AI capabilities locally while keeping data on their devices, enhancing privacy, security and accessibility."

PIB does not give a parameter count for either Dhīr model. Those figures are on zenteiq.ai/dhir, which BharatBeat reviewed for this piece. ZenteiQ describes Dhīr 2B as a dense transformer with 2.03 billion parameters, and Dhīr 5B as a mixture-of-experts model with 5.09 billion parameters in total and 1.79 billion active per token (2 of 12 experts per token). Both are listed with a 16,384-token context, and the site says they are released as instruction-tuned checkpoints that share a tokenizer and training recipe.

The languages named on that page are Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi and Odia, alongside English. That is the company's list of the 10 Indian languages. PIB does not name them.

ZenteiQ says the models were pretrained from scratch on TPUs in India: every training stage ran on Google Cloud TPU v6e in Mumbai, with data and checkpoints held in the same region, and that each model saw more than 6 trillion pretraining tokens. Those are company statements. They are separate from PIB's figure of about 7.5 trillion tokens for EKA-1. The site also says a larger Dhīr 9B is in development, with public release subject to evaluation.

Five released, of 20 backed, according to Moneycontrol

Moneycontrol reported that MeitY on 3 November released two indigenous foundational models under the IndiaAI Mission, by Soket and ZenteiQ. "With this, out of the 20 backed by the IndiaAI Mission, the total number of models released has increased to five."

The other models Moneycontrol says have already been released include:

• BharatGen's Param 2, a foundational model built up to 17 billion parameters.
• Gnan.ai's Vachana, a voice AI and text-to-speech foundational model.
• Sarvam AI's suite of multilingual models for document processing, translation and Indic language understanding.

Moneycontrol's "two" sits next to three names (EKA-1, Dhīr 2B and Dhīr 5B). This piece does not try to reconcile how the mission counts a model family versus a single checkpoint. The five-of-20 line is Moneycontrol's, not a figure restated in the PIB note read for this article.

Where the mission itself comes from

PIB's note on the mission says the Union Cabinet approved IndiaAI on 7 March 2024, guided by the vision of "Making AI in India and Making AI Work for India". It lists seven pillars: IndiaAI Compute Capacity, IndiaAI Innovation Centre, the IndiaAI Datasets Platform (AIKosh), the IndiaAI Application Development Initiative, IndiaAI FutureSkills, IndiaAI Startup Financing, and Safe & Trusted AI.

EKA-1 and Dhīr are presented as products of the Innovation Centre pillar, with AIKosh named as the public route in. Krishnan's autonomy line is about building the models in India and putting them to work in government, industry, research and education. Shrivastava's line is about range: small enough for everyday devices, large enough for enterprise jobs. Neither quote is a procurement order.

What is still missing

The launch record is a PIB note, company remarks, and same-evening coverage. It is not a model card. Missing from PIB: EKA-1's licence, weights location, evaluation table, and a definition of "sparse hybrid". Missing from PIB, and supplied only by ZenteiQ: Dhīr parameter splits, context length, the 10 language names, and the TPU training account. Krishnan said AIKosh can make the models accessible. He did not say a public endpoint was live on 3 November. No independent lab result is cited in the PIB, Moneycontrol or Economic Times reports read here.

What to watch next

1. AIKosh entries for EKA-1 and Dhīr, including who can download weights and on what licence.
2. The rest of the 20 models Moneycontrol says IndiaAI has backed, and whether the released count stays at five or moves.
3. Independent checks of the 26.3 billion and 7.5 trillion figures for EKA-1, and of ZenteiQ's 2.03 billion and 5.09 billion figures.
4. Offline and on-device pilots in the sectors the speakers named: banking and cybersecurity for EKA-1, and defence, government, industry and education for Dhīr.
5. Dhīr 9B, which ZenteiQ's site says is in development, with any public release still subject to the company's evaluation.

MeitY has put Soket's EKA-1 and ZenteiQ's Dhīr pair on the IndiaAI record. The next proof is public access and outside evaluation, not another line in a launch note.

---

This explainer is based on Press Information Bureau release PRID 2332315, posted by PIB Delhi at 10:02 pm on 3 November 2026; Moneycontrol, 3 November 2026 (10:57 pm IST); The Economic Times Online, last updated 3 November 2026 at 11:08 pm IST; and ZenteiQ's Dhīr model page (zenteiq.ai/dhir) for parameter counts, context length, language names and the company's account of TPU pretraining. Company statements on intended use and future models are attributed as such. PIB does not, in that release, publish a download link or an independent evaluation.