EKA-1, DHIR 2B, DHIR 5B: INDIA'S INDIGENOUS AI MODELS EXPLAINED

Why In News?

The Union Government recently launched three indigenous AI foundation models, Eka-1, Dhir 2B and Dhir 5B. They were developed by Indian companies with support under the IndiaAI Mission.

What Is Eka-1?

• Developer: Eka-1 was developed by Soket Labs Technology and Research (Soket AI Labs) under its Project EKA.  

• Size: It is a foundation model with 26.3 billion parameters, trained on about 7.5 trillion tokens.  

• IndiaAI Support: Soket AI Labs was among the first startups selected under the IndiaAI Mission to build sovereign foundation models.  

Key Features

• Indian Languages: It works across all 22 scheduled Indian languages and English.  

• Open And Sparse: It is an open, sparse hybrid language model. "Sparse" means only part of the model is activated for each task, which saves computing power.  

• Advanced Tasks: It is designed for mathematics, coding, reasoning and agentic use, meaning it can plan and carry out multi-step tasks.  

What Are Dhir 2B And Dhir 5B?

• Developer: These two open-source models were developed by ZenteiQ AiTech Innovations, a Bengaluru-based AI startup.  

• Small Language Models: With about 2 billion and 5 billion parameters, they are lightweight compared with large models.  

Key Features

• Languages: They work as secure, multilingual assistants in English and 10 Indian languages.  

• Flexible Hardware: They can run on different computing platforms, including GPUs and TPUs.  

• GPUs (Graphics Processing Units) and TPUs (Tensor Processing Units) are specialized hardware accelerators designed to handle heavy parallel workloads, especially artificial intelligence and deep learning

• On-Device Use: Their lightweight design allows them to run on mobile phones without internet connectivity.  

• On-Premise Deployment: They can run inside an organisation's own systems, keeping sensitive data secure.  

Applications

• Sectors: They support offline, document-based automation, knowledge retrieval and citizen or enterprise services in defence, government, industry and education. 

• Hard Environments: They are especially useful in low-connectivity or sensitive settings, such as border areas or secure facilities.  

What Is The IndiaAI Mission?

• The Union Cabinet approved the IndiaAI Mission in March 2024, with an outlay of about ₹10,372 crore over five years. 

• Implementing Agency: It is implemented by the IndiaAI Independent Business Division under Digital India Corporation, under MeitY.  

Seven Pillars

• IndiaAI Compute Capacity: Affordable access to high-end GPUs for startups and researchers.  

• IndiaAI Innovation Centre: Development of indigenous foundation models.  

• IndiaAI Datasets Platform (AIKosh): Access to quality, non-personal datasets.  

• IndiaAI Application Development Initiative: AI solutions for public problems.  

• IndiaAI FutureSkills: AI education and training.  

• IndiaAI Startup Financing: Funding support for AI startups.  

• Safe and Trusted AI: Responsible AI tools and frameworks.  

Foundation Model Push

• Multiple Startups Selected: Startups such as Sarvam AI, Soket AI Labs, Gnani.ai, Gan.ai and ZenteiQ were selected to build sovereign models.  

Why Are Indigenous AI Models Important?

Strategic Autonomy

• Sovereign AI: Indigenous models reduce dependence on foreign AI systems for critical sectors such as defence and governance. 

• Data Security: On-premise and on-device models keep sensitive data within India.  

Linguistic Inclusion

• Serving All Indians: Most global AI models are trained mainly on English. Indic models can serve people in their own languages.  

• Example: BHASHINI, launched in 2022 under the National Language Translation Mission, offers AI-based translation across Indian languages. Models like Eka-1 can strengthen such language services.  

Digital Inclusion

• Offline Access: Phone-based models like Dhir can bring AI services to remote areas with poor internet.  

What Are The Key Challenges?

Compute And Cost

• Expensive Training: Building large models needs massive computing power and money, where India lags behind the US and China.  

• Example: In January 2025, China's DeepSeek released a powerful model built at a reported fraction of the usual cost. This encouraged India to pursue cost-efficient indigenous models instead of waiting for frontier-scale budgets.  

Gaps

• Few Indic Datasets: High-quality digital text and speech data in many Indian languages is scarce.  

Talent And Funding

• Brain Drain: Many top AI researchers work abroad.  

• Patient Capital: Foundation model research needs long-term investment that Indian startups often lack.  

Safety And Trust

• Bias And Misuse: AI models can produce biased or harmful outputs and need strong safety testing. 

What Is The Way Forward?

Build The Ecosystem

• Expand Compute: Keep scaling subsidised GPU access under the IndiaAI Compute Portal.  

• Open Indic Datasets: Expand AIKosh with quality datasets in all scheduled languages.  

Adoption In Governance

• Public Sector Use: Deploy indigenous models in government services, education and agriculture.  

Learn From Global Models

• Sovereign AI Strategies: Other countries are building national AI champions.  

• Example: France's Mistral AI has built competitive open models with government and private support, showing that mid-sized countries can develop sovereign AI.  

Responsible AI

• Safety Standards: Use the IndiaAI Safety Institute to test models for bias, safety and reliability.  

Conclusion

Eka-1 and the Dhir models show India's growing ability to build AI that speaks its languages and works in its conditions, from data centres to offline phones. Sustained investment in compute, Indic data, talent and safety can turn these early models into a strong foundation for sovereign, inclusive AI.

Source: THEHINDU