India Releases Three Homegrown Foundation Models: EKA-1 and Dhir 2B, 5B

What happened

The release event took place in New Delhi on Nov 3, attended by S. Krishnan, Secretary of India's Ministry of Electronics and Information Technology (MeitY), senior MeitY and IndiaAI officials, and the leadership teams of Soket and ZenteiQ. The three models come from teams selected under the IndiaAI Innovation Centre (Foundation Models) pillar, with development supported by subsidised compute and research support.

EKA-1, developed by Soket Labs, has 26.3 billion parameters and was trained on approximately 7.5 trillion tokens. It uses an open, sparse hybrid architecture, covers all 22 Scheduled Indian languages alongside English, and is designed for mathematics, code, reasoning and agentic use. Soket CEO Abhishek Upperwal said the system was developed with India's banking, financial services and cybersecurity requirements in mind.

Dhir 2B and Dhir 5B, developed by ZenteiQ, are named for roughly 2-billion and 5-billion parameter scales, support English and 10 Indian languages, and target document automation, information retrieval and multilingual assistance across defence, government, industrial and educational settings. ZenteiQ said both models run fully offline on laptops and edge devices, with weights released under the Apache 2.0 licence alongside technical documentation and evaluation tools, and the models are available through AIKosh.

Key facts

1. Release: November 3, 2026, New Delhi, unveiled by MeitY Secretary S. Krishnan, with the three models supported by the IndiaAI Mission
2. EKA-1: developed by Soket Labs, 26.3 billion parameters, trained on approximately 7.5 trillion tokens, an open sparse hybrid language model covering all 22 Scheduled Indian languages alongside English
3. EKA-1 uses: designed for mathematics, code, reasoning and agentic use, developed with banking, financial services and cybersecurity requirements in mind
4. Dhir 2B and Dhir 5B: developed by ZenteiQ, at roughly 2-billion and 5-billion parameter scales, supporting English and 10 Indian languages
5. Dhir deployment: runs fully offline on laptops and edge devices, can be deployed on an organization's own infrastructure, with weights released under the Apache 2.0 licence
6. Dhir settings: document automation, information retrieval and multilingual assistance across defence, government, industrial and educational settings
7. Access: the models are available through AIKosh for researchers, students, innovators, enterprises and the public
8. Selection background, as reported: Soket was selected in May 2025 to develop India's first open-source 120-billion-parameter foundation model, and ZenteiQ was selected in September 2025 to build India's Scientific Foundation Model
9. Government statement: Krishnan said the initiative would strengthen India's strategic autonomy in AI and expand the availability of models suited to domestic requirements