New Delhi: An AI model that needs the cloud to answer a question is of little use inside a defence installation, a district office with patchy broadband or a school with one shared computer. Three models launched in the capital on Wednesday are built for exactly those rooms.
ZenteiQ AiTech Innovations released Dhir 2B and Dhir 5B, and Soket Labs Technology and Research released EKA-1, all under the IndiaAI Mission. The startups say the models support multiple Indian languages and can run offline, with no internet connection at all.
The pitch is document-grounded automation — a model that answers from the files placed in front of it rather than from whatever it saw during training. That covers knowledge retrieval, citizen-facing services and enterprise workflows. The named sectors are defence, government, industry and education, with banking, financial services and cybersecurity also in view.
The offline claim is the one that matters for government buyers. Departments that cannot send documents to a foreign server have so far had two choices: build in-house or go without. A model that runs inside an air-gapped machine changes the procurement question from "which cloud" to "which laptop".
Dhir 2B and 5B refer to parameter counts, the rough measure of a model's size. Smaller models run on modest hardware, which is what makes offline use practical. EKA-1's size and architecture have not been disclosed.
What has not been released either: the licence terms, where the models can be downloaded, benchmark scores in Indian languages, or the training data. Those details decide whether a district office can actually deploy one, and whether a startup can build a product on top.
India's AI push has so far been heavy on announcements and light on shipped products. Three models that work without a connection are a concrete step, and a testable one. The next few weeks will show whether they hold up outside a demo.