IndiaAI Mission Releases Soket's 26.3-Billion-Parameter EKA-1 and ZenteiQ's Offline-Ready Dhir Models

India's government-backed push to build home-grown foundation models added three model releases, reported by Akashvani on 4 November 2026. In New Delhi, the IndiaAI Mission under the Ministry of Electronics and Information Technology (MeitY) unveiled EKA-1, built by Soket AI Labs, and two compact models, Dhīr 2B and Dhīr 5B, built by ZenteiQ AiTech Innovations. MeitY Secretary S. Krishnan attended the launch.

What was released

EKA-1 (Soket AI Labs). Launch reports describe EKA-1 as a 26.3-billion-parameter "open, sparse hybrid" language model trained on about 7.5 trillion tokens. It covers all 22 scheduled Indian languages plus English. Soket says it was built for mathematics, coding, reasoning and agentic workflows. Coverage also says it is aimed at banking, financial services and cybersecurity, sectors where regulators and customers want data to stay in India.

A "sparse" design activates only part of the network for each token. That lets a large model run at the cost of a much smaller one, which matters in a market where inference spending often decides whether an AI product is viable. Actual inference hardware needs depend on model architecture, quantisation and workload; the launch report does not establish smartphone performance.

Dhīr 2B and Dhīr 5B (ZenteiQ). ZenteiQ took a different approach. The launch report describes compact models for offline-capable, document-grounded automation, knowledge retrieval and public or enterprise services. Detailed supported-language and licensing claims should be checked against their model cards. They are designed for on-premises and mobile deployment without cloud access. For government departments, hospitals, banks and defence users that cannot send data to a public cloud, a small model that runs inside their own network can matter more than a frontier-scale one.

What we don't know yet

The launch materials leave several gaps. Reports note that the initial announcement did not include full benchmark evaluations, comparative reasoning scores, context-window limits or detailed licensing terms for EKA-1. No public coverage gives token counts for the Dhīr models. Developers will be able to judge how these models compare with open-weight models from Meta, Alibaba or Mistral, or with Sarvam's Indic models, only when model cards and evaluation suites are published.

This matters because sovereign models compete with free, capable global alternatives. MeitY Secretary Krishnan has said before that any Indian foundation model has to compete globally to succeed. Independent evaluation on Indic benchmarks is the obvious next step.

Context: from compute subsidies to shipped models

The IndiaAI Mission was approved in 2024 with an outlay of roughly ₹10,372 crore over five years. Its largest component subsidises GPU compute for start-ups and researchers. A second pillar funds a small group of teams to build foundation models trained largely on Indian data. Sarvam, BharatGen, Gnani.ai and others have released models under the programme earlier this year. Soket and ZenteiQ have now joined that list.

Why it matters

Language is the most obvious gap the programme is trying to close. Global models still perform unevenly across Indian languages, especially lower-resource ones. A model that handles all 22 scheduled languages, if its quality holds up, could support citizen services, vernacular customer support and education tools that reach people who do not use English online.

The second gap is deployment. Many of India's largest AI users, including public-sector banks, state governments and the armed forces, work under data-localisation rules or air-gapped security requirements. Open models that run on local hardware give them an option that does not depend on foreign cloud APIs. Developers should check the actual model licence before fine-tuning or commercial distribution.

Sources

• https://newsonair.gov.in/govt-launches-3-new-indigenous-ai-foundation-models-under-indiaai-mission-to-strengthen-homegrown-ai-capabilities/
• https://analyticsindiamag.com/ai-news/indiaai-mission-releases-3-indigenous-ai-models-developed-by-soket-zenteiq
• https://egov.eletsonline.com/2026/11/indiaai-mission-launches-eka-1-and-dh-r-foundation-models-for-public/
• https://www.convergence-now.com/artificial-intelligence/meity-releases-two-more-indigenous-ai-models-under-indiaai-mission-what-are-their-features/
• https://letsdatascience.com/news/indiaai-mission-releases-eka-1-and-dhr-language-models-95f0b9f7
• https://egov.eletsonline.com/2026/11/indiaai-s-next-phase-could-be-about-use-not-just-models/
• https://inc42.com/startups/soket-ai-wants-to-turn-the-language-gap-in-ai-into-indias-edge/
• https://theprint.in/economy/budget-2026-allocates-rs-1000-crore-for-indiaai-mission-pushes-data-centres-and-ai-upskilling/2842947/