India's New AI Models: Meet EKA-1, Dhīr 2B and Dhīr 5B, Built for India's Digital Future
India is taking an interesting step towards making this possible.
On November 3, 2026, the Ministry of Electronics and Information Technology (MeitY), under the IndiaAI Mission, announced three indigenous artificial intelligence models: EKA-1, Dhīr 2B and Dhīr 5B at a launch event in New Delhi.
The models have been developed by two Indian technology companies, Soket Labs and ZenteiQ AiTech Innovations, with support from the IndiaAI Mission.
But what makes this announcement particularly interesting is that the three models have different purposes.
EKA-1 is designed for complex mathematical tasks, coding, reasoning and agentic applications, while Dhīr 2B and Dhīr 5B are smaller, open-source multilingual AI models designed to work locally on devices, including in offline environments.
Together, they represent an important development in India's effort to build AI technology around its languages, businesses, infrastructure and security requirements.
For entrepreneurs, technology companies, developers and businesses, this development raises an important question:
Could India's future AI ecosystem become more accessible, affordable and independent of foreign cloud-based platforms?
Let's explore the three models, what makes them different and what their launch could mean for India's digital economy.
1. India's New AI Models: What Are EKA-1, Dhīr 2B and Dhīr 5B?
Quick summary: India has introduced three locally developed AI models with different strengths, ranging from advanced reasoning and coding to lightweight multilingual assistance that can work without internet connectivity.
The three models were introduced under the IndiaAI Mission's Innovation Centre initiative, which supports the development of indigenous foundation models.
Here is a quick overview:
| AI model | Developer | Main purpose |
|---|---|---|
| EKA-1 | Soket Labs | Mathematics, coding, reasoning and agentic AI |
| Dhīr 2B | ZenteiQ | Multilingual, offline AI assistance |
| Dhīr 5B | ZenteiQ | Multilingual, offline AI assistance |
EKA-1: A model designed for advanced reasoning and coding
Developed by Bengaluru-based Soket Labs Technology and Research, EKA-1 is the largest of the three models.
According to the official government announcement, it features:
-
26.3 billion parameters.
-
Training on approximately 7.5 trillion tokens.
-
Support for all 22 scheduled Indian languages and English.
-
Capabilities targeting mathematics, programming, reasoning and agentic applications.
The model is built as an open, sparse hybrid language model.
Its development also focuses on India's banking, financial services and cybersecurity requirements.
This makes EKA-1 particularly interesting for applications that require more than basic question-answering, including technical research and complex enterprise workflows.
Dhīr 2B and Dhīr 5B: Smaller models built for local AI assistance
While EKA-1 focuses on advanced workloads, ZenteiQ's Dhīr models follow a different approach.
Dhīr 2B and Dhīr 5B are open-source multilingual foundation models designed for local deployment.
They support English and 10 Indian languages:
Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi and Odia.
Their primary applications include document-based automation, information retrieval and multilingual assistance.
The most interesting capability is their ability to operate on compatible devices without needing an active internet connection or continuous cloud access.
This could make AI more accessible in environments where internet connectivity, infrastructure availability or data privacy is a concern.
Sources: MeitY's official announcement and India Today.
2. AI Without Internet: Why Dhīr 2B and Dhīr 5B Could Be a Game Changer
Quick summary: On-device AI allows compatible models to process information locally, reducing dependence on cloud servers and creating new possibilities for privacy-conscious and low-connectivity environments.
Most popular AI assistants depend on remote computing infrastructure.
When you ask a question through a conventional cloud-based chatbot, your request is generally sent to a remote server, where the AI model processes it and returns a response.
This requires connectivity and depends on the availability of the service.
Dhīr 2B and Dhīr 5B introduce a different possibility.
Their lightweight architectures are designed to run on local hardware, including compatible laptops, mobile devices and edge computing platforms.
This means users can interact with the AI even when an internet connection is unavailable, provided the model has already been installed and the device has sufficient resources.
Why does offline AI matter?
1. Data privacy
When AI inference happens locally, the information being processed does not need to be sent to an external model-serving cloud for that interaction.
This can be valuable for organizations handling confidential documents, internal records and sensitive business information.
However, local processing alone does not guarantee complete security. Device security, application permissions, data storage and deployment practices still matter.
2. Accessibility in low-connectivity areas
Internet connectivity is not equally reliable everywhere.
An AI model that operates offline could support education, information retrieval and administrative assistance in locations with limited connectivity.
3. Lower dependence on cloud infrastructure
Cloud AI services can involve API charges, network requirements and ongoing infrastructure costs.
Local inference creates an alternative for certain workloads, particularly when organizations can use existing hardware.
4. Greater control over deployment
Businesses and institutions can explore private, on-premise or fully offline deployments depending on their security and operational requirements.
How lightweight are these models?
ZenteiQ's technical documentation provides more detail:
-
Dhīr 2B has approximately 2.03 billion parameters.
-
Dhīr 5B MoE has approximately 5.09 billion total parameters, with around 1.79 billion active per token.
-
Both have a 16,384-token context window.
-
Their quantized model files are approximately 1.21 GB and 2.95 GB respectively.
The developer has also published hardware performance measurements covering laptop integrated graphics, NVIDIA Jetson Orin NX and Raspberry Pi 4.
These specifications explain why the models are designed for edge deployment rather than relying exclusively on large remote data centres.
Technical details: ZenteiQ's Dhīr models.
3. How These Indian AI Models Could Transform Different Industries
Quick summary: Indigenous AI models could support multilingual services, document retrieval, education, industrial workflows and privacy-sensitive applications across India.
The real value of these models will depend on how effectively businesses and institutions use them to solve practical problems.
India has a diverse economy with different languages, industries and technology requirements. A single cloud-based AI solution may not be ideal for every environment.
Here are some potential applications.
Education and regional-language learning
Students in different parts of India may benefit from AI assistance that can communicate in their preferred languages.
For example, a locally deployed AI assistant could help students understand educational documents, retrieve relevant information and access learning support.
Schools and educational institutions could also explore multilingual information systems.
However, educational accuracy and source verification remain important, particularly when dealing with technical subjects.
Healthcare and medical administration
Healthcare organizations work with sensitive patient information and administrative records.
Local AI systems could potentially assist with document retrieval, internal knowledge systems and administrative workflows without sending every request to an external cloud service.
Any medical application would still require appropriate validation, clinical oversight and privacy safeguards. These models should not be treated as independently reliable medical decision-makers.
Manufacturing and industrial operations
Manufacturing businesses could use locally deployed AI assistants to retrieve information from technical manuals, maintenance records and internal process documentation.
Imagine a technician searching a machine maintenance manual through a local AI interface instead of manually searching hundreds of pages.
Such systems could improve access to organizational knowledge, particularly in industrial environments where connectivity is limited or sensitive information must remain on-premise.
Government and public services
Government offices manage large quantities of documents, citizen enquiries and internal information.
Multilingual AI assistants could help retrieve relevant information, support document-based workflows and improve access to public-service information.
This could be especially useful for services designed for citizens who prefer regional languages.
Banking and financial services
EKA-1's intended focus on banking, financial services and cybersecurity makes it relevant to India's growing financial technology ecosystem.
Potential applications include technical assistance, information processing and specialized workflows.
Financial applications must still meet strict data security, accuracy, auditability and regulatory requirements.
The larger opportunity is to develop AI systems that are adapted to India's operational and language requirements rather than relying exclusively on solutions designed for other markets.
4. What Makes Indigenous AI Important for India's Technology Ecosystem?
Quick summary: Homegrown AI development can strengthen technological autonomy, expand local innovation and create more options for businesses that need control over their infrastructure and data.
For a long time, the global AI market has been dominated by large technology companies with substantial computing infrastructure, research capabilities and investment resources.
India has a massive opportunity to participate in this ecosystem not only as a consumer of AI products but also as a developer of foundational technology.
The launch of EKA-1 and Dhīr reflects this broader ambition.
The role of the IndiaAI Mission
The IndiaAI Mission was approved by the Union Cabinet on March 7, 2024.
Operating under MeitY, it aims to develop India's AI ecosystem through several areas, including computing infrastructure, indigenous foundation models, datasets, application development, future skills, startup financing and responsible AI.
Soket Labs and ZenteiQ were selected under the IndiaAI Innovation Centre's foundation-model initiative.
The government provided access to subsidized national AI computing infrastructure and support for additional research expenses.
This is important because developing foundation models requires significant computing resources, specialized engineering capabilities and extensive research.
Public support can help Indian AI companies experiment, develop their technology and compete in a rapidly evolving global market.
Open-source innovation and accessibility
Dhīr 2B and Dhīr 5B have been released under the Apache 2.0 licence, according to ZenteiQ.
This is significant for developers and businesses because it provides a permissive framework for using, modifying and building applications around the model weights, subject to the licence terms.
The models are also available through India's AIKosh platform.
This creates opportunities for developers, researchers, universities and startups to experiment with locally developed AI technology.
The benefit is not limited to the models themselves.
A growing ecosystem of AI developers, infrastructure providers and application builders can create additional opportunities for Indian technology businesses.
5. Limitations, Business Opportunities and What Comes Next for Indian AI
Quick summary: The new models create opportunities, but their real impact will depend on reliability, hardware compatibility, deployment quality and the ability to solve real-world problems.
While the announcement represents progress, it is important to understand that smaller offline models are not direct replacements for every large cloud-based AI system.
Each model has been designed for particular requirements.
What are the limitations?
Dhīr's developer explicitly identifies some limitations:
-
The models are text-only, without image, audio or speech input and output.
-
They do not include a dedicated reasoning mode.
-
Their unsupported factual recall is a weak area.
-
Larger models can outperform them in mathematics and coding.
For high-stakes subjects such as medicine, finance, agriculture and law, ZenteiQ recommends authoritative document retrieval and human review.
This is why businesses should evaluate each model based on the task it is expected to perform rather than assuming that a smaller AI model can replace more advanced systems.
What does this mean for Indian businesses?
For entrepreneurs and organizations, the launch creates several potential opportunities.
AI-powered customer assistance: Businesses could explore multilingual assistants for answering routine customer questions using approved product and service documentation.
Internal knowledge management: Organizations could build local systems that retrieve information from internal files, manuals and operational documents.
Regional-language automation: Companies serving customers across multiple Indian states could explore AI-driven information access in supported regional languages.
Privacy-conscious deployments: Organizations handling sensitive data could evaluate on-premise and offline model deployment instead of relying exclusively on external cloud services.
AI application development: Indian startups and software companies can build specialized applications around indigenous models rather than developing foundation models themselves.
The CodeAgni perspective: Turning AI technology into business value
At CodeAgni, we believe developments like EKA-1 and Dhīr represent an important direction for India's digital economy.
However, technology alone does not create business growth.
The real opportunity is to integrate useful AI capabilities into everyday business processes.
For example, businesses can explore AI-powered customer support, document-based knowledge assistants, lead-management automation, multilingual communication and connected marketing workflows.
Combined with platforms such as n8n and other automation tools, AI can help businesses streamline repetitive activities and improve operational efficiency.
The right model depends on the use case, available hardware, privacy requirements and expected accuracy.
Our approach is to focus on business problems first and select technology based on the value it can create.
Conclusion: India's AI Future Is Becoming More Local, Accessible and Diverse
The launch of EKA-1, Dhīr 2B and Dhīr 5B marks an important development in India's journey towards building indigenous artificial intelligence.
EKA-1 targets more demanding mathematics, coding, reasoning and agentic workloads, while Dhīr introduces a more lightweight approach to multilingual AI that can operate locally without internet connectivity.
Together, these models address different requirements across India's business, technology and public-service ecosystems.
There are still challenges involving model quality, hardware compatibility, deployment and security. But the ability to develop and experiment with AI systems built around Indian languages and local needs is an important step.
For developers, it creates new opportunities to build applications.
For businesses, it creates more options for adopting AI.
And for India, it strengthens the possibility of developing technology that is not just consumed locally but built locally.
The next chapter of India's AI story will not be defined only by how powerful its models become. It will also depend on how useful, accessible and relevant they are to the people and businesses using them.
Frequently Asked Questions
1. What are India's new AI models launched in November 2026?
India's three new indigenous AI models are EKA-1 by Soket Labs and Dhīr 2B and Dhīr 5B by ZenteiQ. They were released under the IndiaAI Mission on November 3, 2026.
2. What is the EKA-1 AI model?
EKA-1 is a 26.3-billion-parameter foundation model developed by Soket Labs. It was trained on approximately 7.5 trillion tokens and is designed for mathematics, coding, reasoning and agentic applications. It supports all 22 scheduled Indian languages and English.
3. Can Dhīr 2B and Dhīr 5B work without the internet?
Yes. Both models are designed for local and offline deployment on compatible devices, including laptops and mobile or edge hardware. They do not require a continuous cloud connection for local inference.
4. Which Indian languages do the Dhīr models support?
Dhīr 2B and Dhīr 5B support Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi and Odia, alongside English.
5. Are Dhīr 2B and Dhīr 5B open source?
Yes. ZenteiQ has released both models under the Apache 2.0 licence, making them available for developers and businesses to explore and build applications around.
6. How is EKA-1 different from Dhīr 2B and Dhīr 5B?
EKA-1 is a larger model designed for mathematics, coding, reasoning and agentic use. The Dhīr models are smaller multilingual assistants optimized for document-grounded automation, information retrieval and offline deployment.
7. Where can developers access India's new AI models?
The IndiaAI Mission has identified AIKosh as an access platform for these models. ZenteiQ also provides model information, technical documentation and download links for Dhīr through its official website.
8. Can Indian businesses use offline AI models for business automation?
Yes, businesses can explore local AI models for document retrieval, internal knowledge systems and multilingual assistance. However, they should evaluate hardware requirements, accuracy, data security and the need for human review before using them in production.
