India has launched its first homegrown foundation models. The IndiaAI Mission, run by the country’s electronics and IT ministry, released three models in New Delhi on 3 November: EKA-1 from Soket Labs, and Dhīr 2B and Dhīr 5B from ZenteiQ. The trio is built for Indian languages and for places where the internet cannot be taken for granted, with two of the models designed to run entirely offline on a phone. It is the clearest step yet in India’s push for AI it owns outright.
## Three models, two very different jobs
EKA-1 is the heavyweight. Soket Labs built the open, sparse hybrid model with 26.3 billion parameters, trained on 7.5 trillion tokens, and tuned it for maths, code, reasoning and agentic work. It covers all 22 scheduled Indian languages alongside English, and Soket pitches it at banking, financial services and cybersecurity, sectors where keeping data in-country matters more than speed.
Dhīr 2B and Dhīr 5B take the opposite approach: small, light and built to work where the internet does not reach. ZenteiQ’s open-source models support English and 10 Indian languages, and run as on-premise assistants for defence, government, industry and education.
The company says compressed versions can run on a phone or laptop with no connection at all, and it plans a free mobile app for Android and iOS. A larger 9-billion-parameter model for researchers and engineers is planned for December.
## The fine print: ‘open’ still means asking first
The launch is real, but the word open needs a footnote. As of 4 November, EKA-1’s page on AIKosh, the model library, listed an Apache 2.0 licence and a 30.49 GB size, yet held only three settings files and no weights. Soket’s site lists the model as coming soon behind a waitlist, with the weights to follow within a month.
Dhīr is further along but not fully open either. Its weight files are on AIKosh, 3.79 GB for the 2B model and 9.50 GB for the 5B, yet all three models are marked Restricted, so downloading them needs the contributor’s explicit approval. Until the weights arrive without a gatekeeper, open here means asking permission, not simply downloading.
## Our read
The substance here is sovereignty, not leaderboards. India is following the same playbook as China, which [wrote full-lifecycle AI rules into law](https://geroundtable.com/china-ndrc-full-lifecycle-ai-law/) this month, and the same logic drives [Europe’s open-weight push](https://geroundtable.com/mistral-large-4-public-open-weight-release/): models a country can inspect, host and adapt without asking a foreign lab.
EKA-1’s 26.3 billion parameters will not trouble [the frontier](https://geroundtable.com/thinking-machines-inkling-2-small/), but a capable model in 22 Indian languages, running in an air-gapped bank or ministry, is a different kind of win. Dhīr’s bet is narrower still: reach the hundreds of millions of Indians whose first, and often only, computer is a phone with patchy signal.
No independent benchmark has been published for any of the three models, and the language counts do not line up: the release says 22 languages, while Soket’s pages quote 60 or more. Until the weights and reports land, this is a launch to watch, not yet a rival to the labs.
## Why It Matters
India is vast, and most of its people do not speak English as a first language. Models that work in 22 Indian languages, on a phone with no signal, put AI into hands the big labs’ products mostly miss. If the weights arrive, India gains something few countries have: models it built, owns and can run on its own soil.
## Background
The IndiaAI Mission was approved by the cabinet in March 2024 and runs across seven pillars, including subsidised national AI compute that helped build these models. Twelve teams received funding to build foundation models, and five have now launched: Sarvam, BharatGen and Gnani at the February AI summit, joined this week by Soket and ZenteiQ.
Source: [IndiaAI Mission release](https://www.pib.gov.in/PressReleseDetailm.aspx?PRID=2332315&lan®=48&lang=2) and [Economic Times report](https://economictimes.indiatimes.com/tech/artificial-intelligence/indiaai-missions-three-indigenous-ai-models-target-offline-mobile-access-secure-use-in-critical-sectors/articleshow/135524432.cms)
0 Comments