EKA-1 and Dhir 2B/5B: IndiaAI's Three Foundation Models

India put three new names on the global AI map this week. On November 3, 2026, the Ministry of Electronics and Information Technology unveiled three indigenous foundation models in New Delhi: EKA-1, built by Bengaluru-based Soket Labs, and Dhir 2B and Dhir 5B, built by the IISc-born startup ZenteiQ. The launch, fronted by MeitY Secretary S Krishnan, is one of the first public outputs of the IndiaAI Mission's foundation-model pillar, the programme through which the government is funding homegrown alternatives to US and Chinese models.

The detail that matters most for developers: the two Dhir models ship with their weights under the Apache 2.0 licence, downloadable through the government's AIKosh platform, and are small enough to run fully offline on laptops and edge devices. EKA-1, the larger Soket Labs model, is aimed at mathematics, coding and reasoning, and it is part of a longer pipeline that includes a planned 120-billion-parameter flagship model. This article breaks down what was actually announced, what remains unconfirmed, and how the three models fit the wider sovereign-AI wave that has already produced Switzerland's Giotto family and Europe's Polarise cloud.

What was announced on November 3

The event itself was a government unveiling, not a product launch with pricing and a download button. According to ANI's report, carried by the Tribune India, MeitY Secretary S Krishnan congratulated the two Bengaluru teams and framed the models as a step toward "strategic autonomy in artificial intelligence" — the government's term for reducing dependence on foreign AI systems. PTI's parallel report, published by ThePrint, confirms the three models and the officials present.

The work sits under the IndiaAI Innovation Centre's Foundation Models pillar, the mission's programme for creating indigenous large language models. Goodreturns, reporting from the event, adds that the teams benefited from subsidised national AI compute and research support, and that the models target Indian-language use cases, including offline, document-grounded services for government and enterprise settings.

Three things were conspicuously absent from every launch report: benchmark scores, API pricing, and for EKA-1, a public download. That is not necessarily a red flag — these are foundation-model releases aimed first at researchers and public-sector builders — but it sets the honest frame for the rest of this article: what follows is what the builders and the government have actually said, and where their accounts disagree, the disagreement is flagged.

The three models, side by side

EKA-1 is the heavyweight of the trio. ANI describes it as a 26-billion-parameter foundation model designed for mathematics, coding and reasoning, trained on 7.5 trillion tokens, and supporting more than 100 Indian and international languages alongside programming languages. That token figure, if accurate, is a serious training run: 7.5 trillion tokens is in the same league as the largest published training budgets for models of this size.

Two caveats belong here, because the launch reporting does not fully agree with itself. First, a detailed Goodreturns breakdown puts EKA-1 at 26.3 billion parameters and says it supports "all 22 scheduled Indian languages and English" — a much narrower language claim than ANI's "more than 100." Second, the CIO/Elets analysis notes that "Soket Labs did not disclose a parameter count for EKA-1 in the launch reports." So the 26B figure comes from agency copy, not from a Soket Labs spec sheet. Treat the size and language numbers as provisional until Soket publishes a model card or technical report.

ZenteiQ's Dhir pair is the better-documented half of the announcement. Per the company's own figures reported by CIO/Elets, Dhir 2B is a dense model with 2.03 billion parameters, while Dhir 5B uses a mixture-of-experts design with 5.09 billion total parameters, of which 1.79 billion are active at a time. Both were pretrained from scratch — not adapted from existing open weights — on Google Cloud TPUs, and both are compact enough, in ZenteiQ's telling, to run fully offline on laptops, edge devices and tested mobile hardware. A larger Dhir 9B is already in development.

Why the Dhir licence matters more than the size

The headline for most readers outside India is not the parameter counts. It is the licence. ZenteiQ is releasing the Dhir weights under Apache 2.0, along with technical documentation and evaluation tools, through AIKosh, the government's AI data and models platform. That makes Dhir one of the few government-backed sovereign models anywhere to ship under a genuinely permissive open-source licence — closer in spirit to Mistral's open-weight releases than to the proprietary national models coming out of other programmes.

Apache 2.0 means anyone, anywhere — including teams in the US, UK and Canada — can download the weights, fine-tune them on their own data, and deploy them commercially, with the standard patent-grant and attribution conditions. For enterprises and public bodies handling data that cannot leave their premises, the pitch is straightforward: a small, multilingual, document-grounded model that runs on your own infrastructure, with no cloud dependency and no foreign-vendor data flow. ZenteiQ's founder, IISc professor Sashikumaar Ganesan, has emphasised that cloud connectivity is not required, and the launch coverage positions the models for defence, government, industrial and education use cases where low connectivity or sensitive data rules out API-based models.

Practical notes before you reach for the download. First, verify the actual licence text on the AIKosh listing when you download — launch reports say Apache 2.0, but the licence file in the repo is what counts. Second, "runs on a laptop" is ZenteiQ's claim about tested hardware; real-world memory requirements will depend on quantisation, and no independent memory benchmarks were published at launch. Third, the evaluation tools shipping with the weights are the company's own — useful, but not a substitute for independent testing on your own data.

How Soket Labs and ZenteiQ got here

The November 3 unveiling is the visible tip of a selection process that began well over a year ago. MeitY selected Soket AI in May 2025 to develop India's first open-source 120-billion-parameter foundation model tuned for the country's linguistic diversity, with defence, healthcare and education as the target sectors. ZenteiQ was named in September 2025 to build India's Scientific Foundation Model, with 75 crore rupees earmarked in its first year, mostly for compute — roughly $7.9 million at November 4, 2026 market rates (about 94.8 INR to the dollar), verified at publication.

Soket Labs is not a newcomer. Founded in 2019 by Abhishek Upperwal, the startup runs Project EKA and earlier presented its EKA Code 24B model to Prime Minister Narendra Modi at the IndiaAI Impact Summit 2026, per the CIO/Elets report. EKA-1 is its first model to carry the IndiaAI Mission's backing into public view, and the planned 120-billion-parameter model remains the flagship commitment — the one that will ultimately be judged against frontier open-weight releases.

ZenteiQ's pedigree is academic: born at the Indian Institute of Science in Bangalore, and led by founder Prof Sashikumaar Ganesan, it builds what it calls scientific foundation models, pretrained from scratch rather than fine-tuned from foreign weights. That "from scratch" claim is worth underlining, because it is what separates a sovereign model programme from a rebranding exercise — though readers should note the training ran on Google Cloud TPUs, so the compute stack still leans on US infrastructure.

The compute backdrop matters too. The CIO/Elets analysis notes the mission is scaling shared compute through a 10,000-GPU tender run by C-DAC, India's advanced-computing agency. Subsidised national compute is the quiet engine behind these launches, and it is the part of the story most relevant to other countries running their own sovereign-AI programmes.

What is still unconfirmed

An honest account needs this section. As of November 4, 2026, no public benchmark scores exist for any of the three models — no math, coding or reasoning evaluations, no multilingual test results. Every performance impression at this stage is the builders' own description. Soket Labs has not published a model card, technical report or weight download for EKA-1, and its licence status is unconfirmed. The parameter-count and language-support figures for EKA-1 conflict across the launch reports, as noted above. There is no pricing because there is nothing to buy: these are research and deployment artefacts, not products with API rate cards. And while the Dhir models are promised on AIKosh under Apache 2.0, availability at the time of the announcement was described in future-tense terms — check the live listing rather than assuming the weights are already mirrored everywhere.

One more naming note for searchers: agency copy variously renders the builder as "Soket Labs," "Soket AI" and "Socket AI," and the models as "Dhir" or "Dhīr." They are the same entities; the variation comes from syndicated copies, not from different companies or models.

Where this fits in the sovereign-AI wave

India is not launching into a vacuum. The CIO/Elets report situates the models alongside CoRover's three BharatGPT models already listed on AIKosh and the government-backed BharatGen sovereign AI programme, and notes Sarvam's push toward a trillion-parameter model as the mission's other flagship bet. Internationally, the pattern is now familiar: Switzerland's Giotto released sovereign models with published API pricing in October, and Europe stood up the Polarise sovereign cloud the same month. India's differentiator is the combination of state funding, a permissive open licence for the small models, and an explicit offline-first design for secure environments.

For readers in the US, UK and Canada, the practical takeaway is the Dhir weights. A permissive-licence, offline-capable, English-plus-Indian-languages model family is a genuinely useful artefact for multilingual document work, edge deployment and research into small-model training from scratch — regardless of what you think of the "strategic autonomy" framing. The open question, and the one worth revisiting in a few months, is whether the benchmarks and the 120-billion-parameter flagship materialise. This is general information, not procurement or legal advice.

FAQ

Who built India's new EKA-1, Dhir 2B and Dhir 5B models?

EKA-1 was built by Soket Labs of Bengaluru, founded in 2019 by Abhishek Upperwal. Dhir 2B and Dhir 5B were built by ZenteiQ, a startup born at the Indian Institute of Science in Bangalore and led by founder Prof Sashikumaar Ganesan. Both were selected under the IndiaAI Mission's foundation-model pillar and unveiled by MeitY Secretary S Krishnan on November 3, 2026.

Can the Dhir models run without an internet connection?

Yes, according to ZenteiQ. The company says Dhir 2B and Dhir 5B run fully offline on laptops, edge devices and tested mobile hardware, with no cloud connectivity required. Enterprises can fine-tune them on their own data and deploy them on their own infrastructure.

Is EKA-1 open source?

Not confirmed. Agency reports describe EKA-1 as a 26-billion-parameter model for math, coding and reasoning, but Soket Labs has not published weights, a model card or a licence for it. The Dhir models are confirmed for Apache 2.0 release via AIKosh; do not assume EKA-1 carries the same terms.

Where can I download the Dhir model weights?

Through AIKosh, the Indian government's AI data and models platform, where ZenteiQ is releasing the weights with documentation and evaluation tools. Check the live listing for the actual licence text and files before building on them.

How do these compare to Sarvam and BharatGen?

They are complementary tracks under the same mission. Sarvam is pursuing a much larger, reportedly trillion-parameter-scale model, and BharatGen is the separate government-backed sovereign AI programme. EKA-1 and the Dhir family are the first public outputs from the mission's foundation-model pillar startups.

What should enterprises watch next?

Three things: published benchmarks for all three models, Soket Labs' planned 120-billion-parameter open-source flagship, and the larger Dhir 9B. CIOs evaluating offline deployment should also watch the mission's 10,000-GPU shared-compute rollout through C-DAC.

India's foundation-model bet is now public, and the Dhir family's Apache 2.0 weights give the rest of the world something concrete to test. The models' real test — benchmarks, deployments and the 120-billion-parameter flagship — is still ahead. If you try the Dhir weights on your own data, compare them against the small-model baselines you already use; sovereign provenance is interesting, but measured quality is what decides adoption.