India Unveils EKA-1 and Dhir AI Models With Indian Language and Offline Capabilities

Three Indian AI models target local languages and private computing

India has unveiled three domestically developed artificial intelligence models under the IndiaAI Mission: Soket Labs' EKA-1 and ZenteiQ's Dhir 2B and Dhir 5B. The November 2026 announcement brings together two priorities for the country's AI programme: software that works in Indian languages and tools that can perform useful tasks without a constant internet connection.

EKA-1 has 26.3 billion parameters and supports English and all 22 languages listed in the Eighth Schedule of India's Constitution. The smaller Dhir models have two billion and five billion parameters respectively and support English and 10 Indian languages. Their developers describe uses ranging from answering questions about documents to helping with mathematics, programming and institutional work.

S. Krishnan, secretary of the Ministry of Electronics and Information Technology, attended the launch with ministry and IndiaAI officials. He said domestically developed models could strengthen India's technological independence and support work in government, industry, research, education and cybersecurity.

The most immediate practical promise is greater choice over where AI runs. An organisation could use a model on its own computers, while some mobile applications could process requests directly on a phone. For institutions handling sensitive documents, this offers a way to use AI while keeping information within their own computing environment.

Access to the models is being made available through AIKosh for researchers, students, startups, companies and the public. ZenteiQ also plans a free mobile application. The announcement therefore covers both the underlying models that developers can build upon and a proposed route for people to use them directly.

What EKA-1 is designed to do

Soket Labs developed EKA-1 for tasks including mathematics, coding and reasoning, meaning working through information to produce an answer. Chief executive Abhishek Upparwal described banking, financial services and cybersecurity as particular areas of focus, with the model designed around the needs of India's digital economy.

A foundation model is an AI system trained broadly enough to support several applications. Developers can build more specialised tools around it, such as a service that answers questions about financial documents or software that helps a programmer write code. EKA-1 is intended to provide that underlying capability rather than serve only one narrowly defined task.

The model's 26.3 billion parameters are the internal numerical values adjusted during training. They help the system recognise patterns and generate responses. Parameter counts indicate the scale of a model, while users also need to consider language quality, response accuracy, speed and the computing resources required for a particular job.

EKA-1's stated training volume is about 7.5 trillion tokens. Tokens are the pieces into which text is divided for processing, sometimes whole words and sometimes parts of words. A second published figure puts the volume at 7.5 lakh crore tokens, equivalent to 75 trillion. That leaves a material discrepancy in the stated training total.

Indian language support has two different scopes

EKA-1's support for English and all 22 scheduled Indian languages is central to its intended domestic role. A broader description also puts its coverage at more than 100 Indian and foreign languages, alongside programming languages. The scheduled language figure identifies its stated Indian language coverage; the larger figure describes a wider range.

Dhir 2B and Dhir 5B are described as supporting English and 10 Indian languages. Their smaller size and narrower stated language coverage give institutions a different option from EKA-1, depending on the documents, languages and equipment involved.

For a student or public employee, language support can determine whether an AI tool is useful in everyday work. A document assistant that accepts the language used in an office or classroom can reduce the need to translate questions into English before asking them. The launch descriptions, however, provide no language by language performance results or comparable accuracy scores for the three models.

Why the Dhir models are smaller

ZenteiQ chief executive Professor Sashikumaar Ganesan described Dhir 2B and Dhir 5B as compact models, explaining their size at the launch:

We have launched two small models. We call them compact models...what we have launched is a two-billion and then five-billion parameter model...

EKA-1 has about 5.3 times as many parameters as Dhir 5B and roughly 13.2 times as many as Dhir 2B. Those comparisons follow from the announced counts of 26.3 billion, five billion and two billion. Smaller models generally require less memory to store their parameters, making them candidates for devices with more limited resources.

ZenteiQ says its models can run directly on mobile devices and across different computing platforms. Ganesan described the intended experience in terms of giving people access without sending each request to a remote service:

We can run these models on our mobile device itself... users can get AI solutions in their hands without even going to the cloud, number one...

The Dhir models are intended to retrieve information from documents, answer questions and support routine work. Defence organisations, government departments, industrial businesses and educational institutions are among the proposed users.

What working without the internet means

An AI service accessed through the cloud normally sends a request to computers elsewhere, which process it and return an answer. Running a model locally moves that processing onto a phone or an organisation's own equipment. Once the necessary software and model are installed, suitable tasks can continue without a live internet connection.

Krishnan highlighted the potential for these models to operate in secure computer networks without internet access. Such networks are relevant to government offices and other sensitive institutions that restrict connections to outside services. A local document assistant could, for example, help staff search internal material while the processing remains inside the organisation.

ZenteiQ's device based approach could also be useful where connectivity is weak or intermittent. People would be less dependent on a reliable connection for every request. EKA-1 is also described as capable of performing some tasks on mobile phones, although the strongest mobile deployment claims concern the compact Dhir models.

Keeping information on a device gives its owner more control over where processing happens. The organisation still has to manage access to files, secure its equipment and decide which documents an application may use. Privacy therefore depends on the way the complete application is deployed, alongside the model's ability to run locally.

Cloud computing helped build models intended for local use

The Dhir models were developed with assistance from Google Cloud tensor processing units, or TPUs. These are specialised processors used for the large volumes of calculations involved in AI. The models are also described as able to operate on systems using graphics processing units, or GPUs.

Training and everyday use are different stages. Training adjusts a model's internal values by processing large amounts of material. Running the finished model to answer a question is called inference. A developer can use powerful cloud equipment during training and later distribute a model that performs inference on local hardware.

The IndiaAI Mission supported both developers with computing resources at reduced rates and assistance for research expenses. That support addresses a major practical cost of creating foundation models: obtaining enough computing capacity to train and develop them.

Krishnan linked the initiative to technological and strategic autonomy. In practical terms, domestic development gives Indian organisations another source of models shaped around local languages and institutional needs, with options for deployment on infrastructure they control.

How people can access the models

AIKosh is the announced access route for researchers, students, startups, businesses and individual users. For developers, access to a foundation model creates an opportunity to build applications around particular documents or tasks. For institutions, the next step is choosing a deployment that suits their equipment, languages and information rules.

ZenteiQ plans to make its mobile application available free through the Play Store and on iOS. Ganesan also said the company intends to introduce a model with nine billion parameters, extending its range beyond Dhir 2B and Dhir 5B.

The announcement gives no release date for the mobile application or the planned nine billion parameter model. It also leaves unspecified the supported phone configurations, memory requirements, download sizes and model licensing terms, details that will matter to people deciding whether they can install or build with the software.

The launch itself was described as taking place on Tuesday, which fell on November 3 in 2026, while November 4 was also given as the unveiling date. Both descriptions place the introduction of the three models in the same week.

What changes for institutions and users

Government departments and businesses now have three additional domestic models to consider for language and document tasks. Their choices differ in scale: EKA-1 combines a larger parameter count with broad stated language coverage, while the Dhir models focus on compact operation, including use on mobile devices.

Banking and cybersecurity are specific targets for Soket Labs. Document retrieval and questions about internal information are prominent uses for Dhir. Schools, researchers and public bodies could also build applications around those capabilities, selecting the model and computing setup that fit their work.

The next concrete developments are wider access through AIKosh, the proposed free mobile application and ZenteiQ's planned larger model. For users, the practical value will come from how well these systems handle their own languages, documents and devices.

Key Points

• India unveiled EKA-1, Dhir 2B and Dhir 5B under the IndiaAI Mission in November 2026.
• EKA-1 has 26.3 billion parameters and supports English and all 22 scheduled Indian languages.
• Dhir 2B and Dhir 5B support English and 10 Indian languages and are designed for compact deployment, including on mobile devices.
• Local processing could support document work in institutions with restricted internet access and in places with limited connectivity.
• The IndiaAI Mission assisted the developers with computing resources and research expenses.
• AIKosh is the announced access platform; a free ZenteiQ mobile application and a model with nine billion parameters are planned.