समाचार Samāchār · Tech & AI · 4 November 2026

IndiaAI releases three home-built models. Now they have to earn users

Soket’s EKA-1 covers all 22 scheduled languages. ZenteiQ’s Dhīr models can run on a phone without the cloud. No independent benchmarks came with them, so use is the test.

India has three more foundation models of its own. The IndiaAI Mission released EKA-1 by Soket Labs and Dhīr 2B and Dhīr 5B by ZenteiQ AiTech Innovations in New Delhi on Tuesday. This is what the programme was built to produce. Indian teams, subsidised national compute, Indian languages.

EKA-1 is the larger bet. It has 26.3 billion parameters and was trained on about 7.5 trillion tokens. It covers all 22 scheduled languages and English. PIB describes it as an open, sparse hybrid language model for mathematics, code, reasoning and agentic use. Soket CEO Abhishek Upperwal said it is “specifically tailored to India’s unique banking, financial services and cybersecurity landscape”.

The Dhīr models go the other way. They are open-source assistants in English and 10 Indian languages, built for on-premise work without continuous cloud access. The target is defence, government, industry and education. ZenteiQ CEO Prof Sashikumar said the lightweight design enables deployment on mobile phones “without dependence on cloud connectivity”. In a country with patchy networks and sensitive data, that is a practical feature.

Credit goes to the teams and to the design of the mission. The Cabinet approved IndiaAI on 7 March 2024. Selected teams received subsidised compute and support for research costs. MeitY Secretary S Krishnan said such models strengthen India’s technological and strategic autonomy in AI. IndiaAI COO Sudeep Shrivastava said the range runs from everyday devices to enterprise systems.

What is still missing is proof. The release and the reports on it give no independent benchmarks. PIB does not state EKA-1’s licence terms or where the models can be downloaded today. Krishnan pointed to AIKosh as the route to researchers, students and the public. The next milestones are public weights, published scores and a first bank or department running one in production.

What a founder takes from it: A sovereign model matters only when it is in use. Founders building for banks, government or low-connectivity India should test these against what they run today and publish the results. An offline model on a phone is the one that can reach the bottom of the pyramid.