Two Indian startups have launched indigenous foundation models under the IndiaAI Mission that work offline and across Indian languages, The Hindu BusinessLine reported. ZenteiQ AiTech Innovations released Dhir 2B and Dhir 5B, while Soket Labs Technology and Research launched EKA-1. The models were unveiled late Tuesday in the presence of MeitY Secretary S Krishnan.
Dhir 2B and Dhir 5B are open-source models built for offline-capable, document-grounded automation and knowledge retrieval in defence, government, industrial and educational settings - particularly low-connectivity or sensitive environments. They can be deployed as secure on-premises assistants supporting English and 10 Indian languages. ZenteiQ CEO Sashikumar said both models run across GPUs and TPUs and work on any smartphone with at least 6GB of RAM, keeping data on the device.
Soket Labs' EKA-1 is the larger effort: a 26.3-billion-parameter foundation model trained on around 7.5 trillion tokens, spanning all 22 scheduled Indian languages alongside English. Built as an open, sparse hybrid language model, it is designed for math, code, reasoning and agentic use. CEO Abhishek Upperwal said the model is tailored to India's banking, financial services and cybersecurity landscape.
The launches extend the IndiaAI Mission's earlier foundation-model releases from Sarvam AI, BharatGen and Gnani AI. Krishnan said indigenous models strengthen India's technological and strategic autonomy, and that platforms such as AIKosh can open them to researchers, students and innovators.
The strategic point is the offline capability. Models that run on a 6GB-RAM smartphone or an air-gapped government server do not need cloud connectivity, foreign data centres or per-token billing - properties that matter for defence outposts, district offices and privacy-sensitive enterprises alike.
FundForge Insight: India's sovereign AI push is splitting into two bets - frontier-scale models and small models that run anywhere. Dhir and EKA-1 are early proof that the second bet can ship. The adoption test is unglamorous: whether government departments and banks actually deploy these models in workflows, not whether benchmarks impress. Watch AIKosh download and deployment patterns over the coming quarters. Sources: The Hindu BusinessLine (November 4, 2026).
