India releases three homegrown AI models for its languages
India unveiled three homegrown AI foundation models on November 3, 2026: EKA-1, a 26.3-billion-parameter model covering all 22 scheduled Indian languages, and the smaller, offline-capable Dhīr 2B and Dhīr 5B. The release is the IndiaAI Mission's bid for AI that India builds and controls itself.
Key takeaways
- On November 3, 2026, India's Ministry of Electronics and IT unveiled three indigenous foundation models in New Delhi: EKA-1 by Soket Labs and Dhīr 2B and Dhīr 5B by ZenteiQ.
- EKA-1 has 26.3 billion parameters, was trained on about 7.5 trillion tokens, and covers all 22 scheduled Indian languages plus English, aimed at math, code, reasoning, and agentic work.
- Dhīr 2B and Dhīr 5B are compact, open-source models for English and 10 Indian languages, built to run offline on laptops, phones, and on-premise systems in low-connectivity or sensitive settings.
- The models are available through AIKosh, the IndiaAI Mission's datasets platform, for researchers, students, enterprises, and the public.
- No benchmark scores or independent evaluations were published with the launch, so how these models compare with global alternatives is still unverified.
What happened
India unveiled three homegrown AI foundation models on November 3, 2026, at a press event in New Delhi: EKA-1, built by Soket Labs, and Dhīr 2B and Dhīr 5B, built by ZenteiQ. The release was presented by S. Krishnan, Secretary of the Ministry of Electronics and Information Technology (MeitY), under the IndiaAI Mission's Innovation Centre pillar for foundation models.
The pitch from the ministry is strategic autonomy. Krishnan said developing indigenous models strengthens India's technological and strategic autonomy in AI, and opens up productivity gains across government, industry, research, and education. Both builders were selected under the mission's foundation-model program, and their work was supported by subsidized national AI compute infrastructure and research funding.
EKA-1: the big model, tuned for India
EKA-1 is the heavyweight of the three. 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 listed in India's Eighth Schedule alongside English, and is designed for math, code, reasoning, and agentic work.
Soket CEO Abhishek Upperwal said the model was tailored to India's banking, financial services, and cybersecurity landscape, built to handle sector-specific requirements and the country's evolving digital ecosystem. That is the target market, not a deployment claim: the launch announcement does not establish that any bank or cyber-defence system is already using EKA-1 in production.
Dhīr: small models that run without the cloud
The Dhīr pair takes the opposite approach. ZenteiQ, a startup born at the Indian Institute of Science in Bengaluru, built Dhīr 2B and Dhīr 5B as open-source models trained from scratch, rather than adapted from existing weights. Dhīr 2B is a dense model with 2.03 billion parameters; Dhīr 5B uses a mixture-of-experts design with 5.09 billion total parameters, of which 1.79 billion are active at a time.
Both support English and 10 Indian languages, and both are built to run offline: on laptops, on phones, and on an organization's own servers. ZenteiQ CEO Prof. Sashikumaar Ganesan said the lightweight design lets the models run on-device on mobile phones without cloud connectivity, keeping data on the device. The intended settings are exactly where a cloud call is a problem: defence, government, industrial, and education work in low-connectivity or sensitive environments. ZenteiQ is releasing the model weights under the Apache 2.0 licence, with documentation and evaluation tools, through AIKosh.
What it means for you
If you build for Indian users, the practical news is language coverage plus deployment choice. A model that handles all 22 scheduled languages natively is aimed at the long tail of Indian-language services that global models serve unevenly, from customer support to citizen services. The Dhīr models matter differently: an offline, on-premise assistant changes the compliance conversation for banks, hospitals, and government offices that cannot send sensitive data to a foreign cloud.
For everyone else, this is a sovereignty play to watch. India is following the same logic as other governments building domestic models: reduce dependence on technology built and controlled elsewhere, and keep the data and the decisions closer to home. The IndiaAI Mission, approved by the Union Cabinet in March 2024, runs seven pillars spanning compute, datasets, skills, and startup financing; these three models are among its first public foundation-model outputs.
What is still unclear: how EKA-1 and the Dhīr models score on public benchmarks, when wider enterprise deployments will appear, what the larger follow-ups look like (Soket is also building a 120-billion-parameter open-source model for Indian languages, and a Dhīr 9B is reportedly in development), and whether the open weights attract a developer community or stay mostly a government showcase. The announcement is real and the models are available; the verdict on their quality is still to come.
Comments