What happened
The Indian Ministry of Electronics and Information Technology (MeitY), through the IndiaAI Mission, has launched three indigenous foundation models: EKA-1, a 26.3-billion parameter sparse hybrid model, and the Dhīr family (5B and 2B parameters). Developed by Soket Labs and ZenteiQ AiTech Innovations respectively, these models were created using state-subsidized compute resources to address the 'Indic Token Tax'—a phenomenon where Western models fragment Indian scripts into excessive tokens, increasing costs and latency. EKA-1 is designed for enterprise and agentic workflows, while the Dhīr models are optimized for offline, on-device execution.
EKA-1 is a 26.3-billion parameter sparse hybrid model developed by Soket Labs. It utilizes a Mixture-of-Experts (MoE) architecture and was trained on 7.5 trillion tokens. It features a 64,000-token and is designed to interface with the Loop agentic framework for autonomous tasks in banking and cybersecurity.
The Dhīr 5B and 2B models, developed by ZenteiQ AiTech Innovations, are small language models (SLMs) designed for 100% offline, on-device execution. They are released under the Apache 2.0 license, allowing for free commercial and academic use.
The models were developed using subsidized access to high-performance computing, including Google Cloud TPU infrastructure and C-DAC supercomputing clusters, as part of the ₹10,372-crore IndiaAI Mission.
A key technical achievement is the development of multilingual tokenizers that support all 22 scheduled Indian languages. This addresses the 'Indic Token Tax,' where Western models often require 4 to 7 tokens for a single Indic word, whereas these models achieve near 1:1 parity, reducing generation latency by up to 75%.
Source details: paidfreedroid.com
Why it matters
This launch represents a strategic shift for India from a consumer of foreign AI to a developer of sovereign foundation models. By engineering proprietary tokenizers that achieve near 1:1 token-to-word parity for 22 Indian languages, these models significantly reduce computational overhead and latency compared to Western alternatives. The release of the Dhīr models under an Apache 2.0 license provides a critical alternative to cloud-tethered, subscription-based AI, enabling secure, offline, and privacy-focused applications in sectors like defense, banking, and mobile computing where data sovereignty is paramount.
The models provide a hedge against 'technological vassalage,' where domestic infrastructure relies on foreign hyperscalers. By controlling the model weights and tokenization, India gains independence from external API pricing and potential geopolitical trade controls on AI compute.
The Dhīr models enable 'zero-cloud' inferencing, which is critical for sensitive sectors like defense and banking. By running locally, these models eliminate the need to transmit confidential data to remote servers, mitigating risks of international surveillance and data breaches.
The release offers a cost-effective alternative for local developers who have historically been limited to fine-tuning foreign open-weight models. The Apache 2.0 licensing removes revenue-share barriers, potentially accelerating the development of a domestic AI application ecosystem.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
The primary challenge for these models will be long-term maintenance and ecosystem adoption. While the initial release provides a technical foundation, the models must demonstrate sustained performance parity with rapidly evolving global frontier models. Observers should monitor whether Indian enterprises transition from foreign APIs to these sovereign models, the integration of these models into government-led digital public infrastructure, and the ability of the developers to maintain competitive update cycles without continued state-subsidized compute access.
The sustainability of the IndiaAI Mission's compute subsidies is a key unknown. It remains to be seen if these startups can maintain development momentum once they move beyond initial state-backed support.
Market adoption will be the ultimate test. The success of EKA-1 and Dhīr depends on whether they can achieve widespread integration into Indian government services and private enterprise workflows, effectively displacing established foreign models.
The developers must prove they can keep pace with the rapid innovation cycles of global frontier models, particularly as generative reasoning becomes the standard for information retrieval and digital services.