On 3 November 2026, in New Delhi, India unveiled three homegrown artificial intelligence foundation models under the IndiaAI Mission — and one of them is a machine built to reason. Soket AI’s EKA-1 is a 26.3-billion-parameter foundation model trained on approximately 7.5 trillion tokens, designed for mathematics, computer programming, reasoning and agentic work, in all 22 languages listed in the Eighth Schedule of the Constitution alongside English.
An illustration of Soket AI’s EKA-1 — a foundation model built to reason through maths, code and multi-step tasks in India’s own languages.
A model built for work where a wrong answer has a cost
The company describes EKA-1 as an open, sparse hybrid language model. The design goal, the company says, is a model that can handle multi-step tasks, execute code and operate inside an organisation’s own walls — on-premises, in a private cloud, or in air-gapped environments where data never leaves the building. The company says EKA-1 can even handle some inference on local devices, so certain uses keep working without mobile data or Wi-Fi.
That shape of deployment is the point. Soket AI’s chief executive Abhishek Upperwal said the system had been developed with India’s banking, financial services and cybersecurity requirements in mind — sectors where answers must be traceable, data must stay inside the perimeter, and a model must show its working rather than simply produce an answer.
Twenty-two languages, one reasoning engine
EKA-1 supports all 22 scheduled Indian languages alongside English, and the company says it also works across programming languages. For a country where much of the most consequential work — a loan file, a compliance check, a classroom lesson, a government form — happens in languages other than English, a reasoning model that works natively in those languages is infrastructure, not a novelty. The model is designed, the company says, to carry a task through its steps: read the problem, work the arithmetic, write and run the code, and arrive at an answer that can be checked.
From a 2019 research bet to a national launch
Soket AI was founded in 2019 by Abhishek Upperwal, and its main model-building effort runs under the name Project EKA. In May 2025, the Ministry of Electronics and Information Technology selected Soket AI under the IndiaAI Mission to build India’s first open-source 120-billion-parameter foundation model tuned to the country’s linguistic diversity, with defence, healthcare and education among its target sectors. Earlier in 2026, the company presented its EKA Code 24B model to Prime Minister Narendra Modi at the IndiaAI Impact Summit — and at Cypher 2026 in Bengaluru in October, Upperwal announced that a 26-billion-parameter model was on its way to open release. EKA-1 is that model, arriving on schedule.
More Read
The models were unveiled in New Delhi by MeitY Secretary S. Krishnan in the presence of senior ministry and IndiaAI officials, and they are accessible through AIKosh, the government’s AI resource platform, for researchers, students, innovators, enterprises and the general public. The release came through the IndiaAI Mission’s Innovation Centre foundation-models pillar, which supports domestic developers building foundation models and large language models — part of a national push to build AI capability at home rather than rent it from abroad.
The stack around the model
EKA-1 does not arrive alone. It sits inside a programme at Soket AI that already includes LOOP, the company’s open-source harness for long-running AI agents, unveiled at Cypher 2026 — the layer that lets an agent stay on a task for hours, days and weeks. A model that can reason, paired with a harness that can keep working, is the combination the company is betting on: agents that derive their answers, show their steps, and run under their owner’s control.
From a research lab that began with a bet on Indian languages, Soket AI has now put a reasoning foundation model into the country’s hands. EKA-1 speaks India’s languages, does its maths in the open, and runs where its owners need it to run. The models get the headlines; this one is built to get to work.
Keep reading: Built From Scratch, Running Offline: Inside ZenteiQ’s Dhīr Models for Indian Engineering · The Gate Before Send: Inside Defmate Secure, Gurugram’s Shield for AI · More Industrial Automation & Technology stories
Like stories like this? Add Exceptional Technology as a preferred source on Google to see more of our reporting in your search results and Discover feed.
