IndiaAI Mission Releases Three Homegrown AI Models: EKA-1, Dhīr 2B and Dhīr 5B
The IndiaAI Mission, under the Ministry of Electronics and Information Technology (MeitY), released three indigenous foundation models (large, general-purpose AI models that other tools can be built on): EKA-1, Dhīr 2B and Dhīr 5B.
EKA-1 was built by Soket Labs Technology and Research. It has 26.3 billion parameters and was trained on about 7.5 trillion tokens. It covers all 22 scheduled Indian languages plus English, and is aimed at banking, financial services and cybersecurity. It is described as an open, sparse hybrid model for maths, code, reasoning and agentic (task-doing) use, and can do some work directly on a phone without internet.
Dhīr 2B and Dhīr 5B (about 2 billion and 5 billion parameters) were built by ZenteiQ AiTech Innovations. They are open-source and support English and 10 Indian languages. They are light enough to run on-device or on-premise (on an organisation's own computers), for use in defence, government, industry and education, especially where internet is weak or data is sensitive.
Both firms were selected under the mission's IndiaAI Innovation Centre (Foundation Models) pillar. The government said subsidised national AI compute (shared GPU computing power) and support for research costs made the models possible.
The models are to be made available through AIKosh, the mission's national platform for datasets and models.
IndiaAI Mission: India's National AI Programme
The IndiaAI Mission is the Government of India's main programme to build a complete AI system inside the country: computing power, datasets, homegrown AI models, skilled people, start-ups and safety tools. It aims to reduce India's dependence on foreign AI and make AI useful for Indian needs and languages. The Union Cabinet approved it in March 2024.
EKA-1 and the two Dhīr models are direct outputs of the mission's foundation models pillar, built using its subsidised compute. Their release on AIKosh shows how the mission's pillars (compute, models, datasets platform) are meant to work together.
Large Language Models (LLMs): How AI Foundation Models Work
A Large Language Model, or LLM, is a computer program that has learned to read and write human language by studying a huge amount of text. It works by predicting the next word (more exactly, the next small piece of a word) again and again, very fast. Because it has seen so much text, this simple trick lets it answer questions, write essays, translate, summarise and even write computer code. Chatbots like ChatGPT are built on LLMs.
EKA-1 is a mid-sized sparse hybrid LLM trained on 7.5 trillion tokens across 22 Indian languages and English. Dhīr 2B and 5B are small language models built to run offline on phones and on-premise systems. Together they show India building both kinds of model: larger ones for complex reasoning, and smaller ones for privacy-sensitive and low-connectivity uses.
- Models released: EKA-1 (Soket Labs Technology and Research), Dhīr 2B and Dhīr 5B (ZenteiQ AiTech Innovations)
- EKA-1: 26.3 billion parameters, trained on about 7.5 trillion tokens, 22 scheduled languages + English; open, sparse hybrid; focus on banking, financial services, cybersecurity
- Dhīr 2B / 5B: open-source, English + 10 Indian languages, on-device and on-premise deployment; runs on GPUs and TPUs
- Released under the IndiaAI Innovation Centre (Foundation Models) pillar of the IndiaAI Mission (MeitY)
- Models to be accessible via AIKosh
- IndiaAI Mission outlay: ₹10,371.92 crore (approved March 2024)