On November 3, 2026, India’s IndiaAI Mission unveiled three homegrown foundation models in New Delhi: EKA-1 from Soket Labs, and Dhīr 2B and Dhīr 5B from ZenteiQ. The release, presented by Ministry of Electronics and Information Technology (MeitY) Secretary S. Krishnan alongside senior MeitY and IndiaAI officials, is a deliberate statement about technological autonomy: models built in India, for Indian languages, and for settings where sending data to a distant cloud is not always desirable or possible. The three models are not three versions of the same idea. They represent two opposite ends of modern AI design — one large, enterprise-focused model and two small, portable ones — and that contrast is what makes the launch interesting.
What IndiaAI announced on November 3
The models were developed under the IndiaAI Innovation Centre (Foundation Models) pillar of the IndiaAI Mission, which supports domestic teams building foundation models with subsidised national AI compute infrastructure. IndiaAI officials framed the release as progress toward making AI in India and making AI work for India, with the models to be made accessible through platforms such as AIKosh for researchers, students and innovators. That public-compute support matters: training a competitive foundation model is expensive, and shared infrastructure lowers the barrier for teams that are not backed by a global hyperscaler.
EKA-1: a larger model tuned for banking and cybersecurity
EKA-1 is the heavyweight of the trio. Soket Labs describes it as an open, sparse hybrid language model with 26.3 billion parameters, trained on approximately 7.5 trillion tokens. It covers all 22 languages in the Eighth Schedule of the Indian Constitution alongside English, and is designed for mathematics, code, reasoning and agentic tasks. Soket CEO Abhishek Upperwal said the model was tailored to India’s banking, financial services and cybersecurity landscape, aiming to understand sector-specific requirements and the country’s evolving digital ecosystem rather than simply translating a general-purpose model into Indian languages. For banks and insurers, that could mean document analysis, compliance support and internal security workflows built around Indian regulatory language and multilingual customer communication — though, importantly, the launch announcement does not establish that EKA-1 has already been adopted by banks or deployed in operational cyber-defence systems.
That enterprise context is also why controls matter: see AI guardrails for enterprise firewalls, which explains why enterprises want controls around AI — context for why enterprises want controls around AI, not a claim about EKA-1 itself.
Dhīr 2B and 5B: small, offline-capable models for sensitive settings
ZenteiQ’s Dhīr models take the opposite approach. Dhīr 2B and Dhīr 5B are open-source, lightweight multilingual assistants supporting English and 10 Indian languages, designed to be deployed securely on-premise and to work offline. ZenteiQ CEO Prof. Sashikumar said they can run across GPUs and TPUs and on mobile phones without dependence on cloud connectivity, keeping data on the device. That design fits defence, government, industrial and educational settings, and places with low or unreliable connectivity, where document-grounded automation, knowledge retrieval and citizen services still need to work. Small models also change the cost equation: they can run on hardware an organisation already owns, rather than renting large-scale inference for every query.
Related reading: secure edge AI connectivity and on-device AI chips running under 3 watts.
Why two very different designs make sense together
Together, EKA-1 and Dhīr sketch a two-track national strategy. Large models handle complex reasoning, long documents and agentic workflows where accuracy and breadth matter most. Small models handle privacy-sensitive, low-connectivity and cost-sensitive work where the model must live close to the user. Many countries are converging on a similar split, but India’s version is distinctive because of language: a model that truly covers 22 scheduled languages plus English is not a localisation afterthought, it is a core design requirement for government, banking and education use. The same logic applies to AI in schools, as seen in Google’s Guided Learning rollout in schools, where language, controls and local context decide whether a tool is actually usable.
What is still unproven — benchmarks, deployment and access
Readers should keep the launch in proportion. The government has not supplied comprehensive independent benchmark results showing how these three models compare with established international alternatives, and detailed deployment figures are not yet available. Parameter counts and training-token totals are useful context, not proof of quality: a 26.3-billion-parameter model can still lag a smaller, better-tuned model on a specific task, and offline capability says nothing about accuracy in a particular language. The honest test will come when researchers and enterprises can try the models on AIKosh, publish independent evaluations, and report where they help and where they fall short.
What to watch next
Watch three things. First, access: how quickly the models, documentation and evaluation tools reach researchers and developers beyond the launch event. Second, real deployments: named pilots in banking, government or education, with measured results rather than launch-day claims. Third, language quality: whether performance in lower-resource Indian languages holds up in practice, not just in a supported-languages list. If those three arrive, EKA-1 and Dhīr will be more than a symbolic launch. They will be the start of an Indian AI stack that other countries with many languages and uneven connectivity will study closely.
Sources: PIB release as reproduced by Technology For You; The 420 report on the launch; CIO / Elets coverage.
