01.Three Indian models, one launch
On 3 November 2026 the IndiaAI Mission, which sits under the Ministry of Electronics and Information Technology, released three foundation models built in India. EKA-1 comes from Soket Labs Technology and Research Private Limited. The other two, Dhir 2B and Dhir 5B, come from ZenteiQ AiTech Innovations. The release was announced at a press event in New Delhi in the presence of MeitY Secretary S. Krishnan and senior officials of the ministry and IndiaAI, along with the leadership teams of both companies.
Both firms were selected under the IndiaAI Innovation Centre pillar of the mission, which is the track meant to build indigenous foundation models and large language models. The press release states that access to subsidised national AI compute infrastructure and support for ancillary research expenses made the work possible. In plain terms, the government paid for a large part of the computing power these two startups needed to train their models.
02.EKA-1 and its 22 languages
EKA-1 is the bigger of the two efforts. It has 26.3 billion parameters and was trained on roughly 7.5 trillion tokens, covering all 22 scheduled Indian languages along with English. The release describes it as an open, sparse hybrid language model built for maths, code, reasoning and agentic use. Agentic here means the model can be wired into software that takes a series of steps on its own rather than only answering a single question.
Soket's stated focus is narrower than a general chatbot. Abhishek Upperwal, the company's CEO, said the model was developed specifically for India's banking, financial services and cybersecurity landscape, and is designed to understand India's evolving digital ecosystem. That framing matters, because banks and financial firms here work under data localisation rules and deal with documents, forms and regulations that are often written in Indian languages or in a very Indian kind of English.
03.Dhir: the small models that work without internet
Dhir 2B and Dhir 5B take the opposite approach on size. At 2 billion and 5 billion parameters they are small enough to sit on modest hardware, and they are open source. The release describes them as secure, on-premise multilingual assistants supporting English and 10 Indian languages, aimed at offline-capable, document-grounded automation and knowledge retrieval. The target users are defence, government, industrial and educational setups, particularly in low-connectivity or sensitive environments.
Prof. Sashikumar, CEO of ZenteiQ, said the models can run across different computing platforms including GPUs and TPUs, and that the lightweight architecture allows deployment directly on mobile phones without depending on cloud connectivity. He said this lets users access AI locally while keeping data on their own devices. For a district office with patchy broadband, or a defence unit that cannot send documents out, that is a different proposition from a cloud subscription.
04.What it means for Indian readers
For students and researchers, the practical entry point is AIKosh, the IndiaAI platform that hosts such resources. S. Krishnan said in the release that through platforms like AIKosh these models can be made accessible to researchers, students, innovators and the wider public. A student building a project in Tamil, Bengali or Marathi has usually had to translate everything into English, run a foreign model, and translate back. A model trained across all 22 scheduled languages removes at least one of those steps.
For small businesses, the Dhir models are the interesting part. A two billion parameter model running on a phone or a single modest server means a clinic, a chartered accountant's office or a wholesale trader can search and summarise their own documents without a monthly cloud bill, and without customer records leaving the premises. For professionals already inside banking, insurance, defence or government, the value is a model that can be installed inside a restricted network instead of being rented from outside the country.
05.The fine print
A launch is not proof of quality. The official release gives parameter counts, training volume and intended uses, but no independent benchmark results, so how EKA-1 performs against established international models on real tasks is not yet publicly known. Sudeep Shrivastava, COO of IndiaAI, framed the release as evidence of progress at both ends of the range, from lightweight everyday models to large enterprise-grade ones, which is a fair description of the ambition rather than a measured result.
The more useful test will come over the next year, and it is about adoption rather than announcements. India has plenty of people who use AI tools and, until now, very few models trained here. Whether colleges, startups and government departments actually build on these three, and whether the compute and support behind them continues, will decide if this turns into a habit or stays a milestone in a press release.
Frequently asked questions
Who built the new Indian AI models?
EKA-1 was built by Soket Labs Technology and Research Private Limited, and Dhir 2B and Dhir 5B by ZenteiQ AiTech Innovations. Both were selected under the IndiaAI Mission's foundation model track and used subsidised national AI compute.
Can these models run without internet?
Yes. ZenteiQ says Dhir 2B and Dhir 5B are lightweight enough to run on mobile phones on-device, without cloud connectivity, so the data stays on the device. They are aimed at low-connectivity and sensitive environments.
How can a student get access to them?
The models are meant to be reachable through AIKosh, the IndiaAI Mission platform for researchers, students and innovators. The official release points to AIKosh as the route for public access and experimentation.