Key Details
The IndiaAI Mission has released models developed by two teams selected under its IndiaAI Innovation Centre pillar. They are designed for different computing and service needs.
Model | Reported Capability and Intended Use |
|---|---|
EKA-1, developed by Soket Labs | A 26.3-billion-parameter model trained on approximately 7.5 trillion tokens. It supports 22 scheduled Indian languages and English and is designed for mathematics, coding, reasoning and AI-assisted tasks. |
Dhīr 2B and Dhīr 5B, developed by ZenteiQ | Open-source models supporting 10 Indian languages and English. They are designed for local, offline-capable use, including work with documents in government, education and other sectors. |
Mission support | The teams received access to subsidised AI computing capacity and support for related research expenses. |
Two Model Families Address Different Needs
A foundation model is an AI system trained on broad data that developers can adapt for tasks such as answering questions, writing code or working with documents. EKA-1 is the larger multilingual model, intended for more demanding tasks. Dhīr 2B and Dhīr 5B emphasise local use: the developer says they can run without a continuous cloud connection, including on suitable devices.
That distinction matters where connectivity is unreliable or an organisation wants to keep data within its own systems. It does not establish that the Dhīr models have already met the security requirements of a defence agency or public department. The release describes intended applications, not results from independent deployment tests.
Public Compute Support Helped Build Domestic Models
The release also shows what the IndiaAI Mission’s investment in shared computing resources is meant to enable: Indian teams developing models for the country’s languages and sectoral needs. MeitY says platforms such as AIKosh could broaden access for researchers and innovators; it does not specify in this release when or on what terms each model will be available through that platform.
Policy Relevance
For public bodies considering these models, language coverage and offline capability are starting points, not measures of service quality. Evaluation would need to examine accuracy across languages, performance on the intended task, data handling and the ability to work reliably on the proposed hardware — particularly before use in services where people depend on correct answers.
The release marks a model-development milestone. Its public value will depend on whether developers can build useful applications from the models and whether those applications work for speakers of the languages they claim to support.
Relevant Question for Policy Stakeholders: What evidence should a public agency require before using an indigenous AI model to answer citizens’ questions or process sensitive documents?
Follow the Full Press Release Here: IndiaAI Mission Releases EKA-1 and Dhīr Models