I noted three homegrown AI models released under the IndiaAI Mission on 3 November. I framed them as designed for India's languages and constraints, and left open whether one large national model or many small offline models matters more.
EKA-1 from Soket Labs is an open 26.3B-parameter model trained on around 7.5 trillion tokens across all 22 scheduled Indian languages plus English.
EKA-1 is designed for maths, code, reasoning and agentic work, and tailored to banking, financial services and cybersecurity.
ZenteiQ's Dhīr 2B and 5B are small open-source assistants that run on-premise and offline, in English and 10 Indian languages.
Dhīr is intended for document-grounded work in defence, government and education.
EKA-1 was not downloadable yet; it was available by waitlist, with open weights planned within a month.
Dhīr's weights need contributor approval, and no independent benchmarks are published yet.
I asked which matters more: one large national model, or many small offline ones?