The IndiaAI Mission released three homegrown models on 3 November. The interesting part is not that India built them, but where they run: on a phone, offline, with no data leaving the device. That quietly changes what "sovereign AI" means, and it gives low-connectivity public services, air-gapped defence work and my own Digital Avatar a realistic substrate. I congratulate Abhishek Upperwal, as I did in June 2025, and ask for the three things the launch did not give us: benchmarks, pricing and deployment details.
Six gigabytes of memory, and no network at all. That is the part of Tuesday's news I cannot get out of my head. Sashikumar Ganesan of ZenteiQ said the Dhīr 2B and Dhīr 5B models run on a mobile phone with at least 6 GB of RAM, that inferencing happens locally, and that "No data will go outside" (Hindustan Times, 4 November 2026).
The IndiaAI Mission released three indigenous foundation models in New Delhi on 3 November, with IT Secretary S Krishnan present (PTI, The Hindu BusinessLine, 4 November 2026). Soket's EKA-1 has 26.3 billion parameters, was trained on about 7.5 trillion tokens, covers all 22 scheduled Indian languages plus English, and leans towards banking, financial services and cybersecurity. It can also do certain inferencing on the local device — a smartphone — with no mobile data or Wi-Fi.
That clause is the story, not the parameter count. For three years "sovereign AI" has meant a big model in a government-funded data centre on subsidised GPUs, answering through an API. On Tuesday the picture inverted: the sovereign model is the one in your pocket, and its sovereignty is proved by nothing leaving it. Sarvam's Prism, released the same day, makes the same move for enterprises — open weights, so a company can retrain the base on its own data and run it on its own iron, though local deployment needs a GPU with 120 GB of memory. Inc42 adds ConvAI's Laya, runnable on laptops and phones, and Sieve from PurpleTalk's Sthānika lab. A cluster, not a one-off.
Three things become realistic that were not realistic last week:
- Low-connectivity public services — an entitlements or health assistant answering in the local language on a mid-range phone, in a block where the tower is unreliable.
- Defence and air-gapped work — Ganesan named defence and cybersecurity; Krishnan said smaller models can be deployed in air-gapped environments.
- My Digital Avatar — every question put to it today travels to a server and comes back. A 2-billion-parameter model on the device would answer with the archive never leaving the phone.
I described that archive in July 2022, when 30,000 documents had become 82,496 memory blocks. The round trip was invisible then, because it was not yet a choice.
What I do not have is the arithmetic. "The announcement did not provide benchmark results, deployment details or pricing information," Inc42 reported. Sarvam published its numbers the same day: 71.8 average across three indices at a 44 ms median on a B200. Soket and ZenteiQ published none. So two questions remain open — how does EKA-1 score on Indian-language and coding benchmarks, and what will an on-device licence cost a state government, on which certified handsets?
One more, gently. The June 2025 report on Soket's selection spoke of an open-source 120-billion-parameter model; EKA-1 arrives at 26.3 billion. A model that ships beats a parameter count, but is the larger one still coming?
Krishnan noted that AIKosh can make these models reachable, a repository I have followed since June 2025. And Abhishek — I congratulated you on 3 June 2025 and I meant it. I mean it more now that something runs. Send the benchmarks, the price and the deployment list, and 3 November stops being a launch and becomes a foundation.
With regards,
Hemen Parekh
www.Iam-Immortal.ai / www.IndiaAGI.ai / www.YourContentCreator.in
04 November 2026
No comments:
Post a Comment