The Indian Ministry of Electronics and Information Technology has launched three new sovereign AI models under the India AI Mission.

The “IndiaAI Mission” was approved by the Union Cabinet of India in March 2024 and is being implemented by the Ministry of Electronics and Information Technology (MeitY). The mission aims to “develop AI in India and make AI work for India.” MeitY has now unveiled three new indigenous AI models developed by two AI laboratories: Soket Labs Technology and Research and ZenteiQ AiTech Innovations. Soket AI claims that its model demonstrates “advanced capabilities” in mathematics, programming, reasoning, and agentic applications. Meanwhile, ZenteiQ states that its open-source large language model (LLM) supports 10 Indian languages.

MeitY Launches EKA-1, Dhīr 2B, and Dhīr 5B Sovereign AI Models
In a press release issued on Wednesday, the Ministry of Electronics and Information Technology (MeitY) announced the launch of three new indigenous AI models under the flagship “IndiaAI” initiative. These models include EKA-1, developed by Soket AI, and Dhīr 2B and Dhīr 5B, developed by ZenteiQ. Soket claims that EKA-1 features 26.3 billion parameters.

Furthermore, this new indigenous AI model was reportedly trained on approximately 7.5 trillion tokens. EKA-1 underwent training using 550 million supervised fine-tuning (SFT) data samples drawn from various domains. It utilizes a hybrid architecture developed by Soket called RALSNet (Recursive Attention Latent Sparse Network), which is claimed to deliver three times the throughput of models of a similar scale. The AI ​​model also incorporates an in-house data pipeline designed for tasks such as web scraping, OCR processing, data labeling and filtering, multi-stage deduplication, tokenization, and data mixing.

Meanwhile, ZenteiQ’s Dhīr 2B and Dhīr 5B are open-source AI models designed for deployment on edge devices to serve as multilingual assistants. These models support English and 10 regional Indian languages. Both Dhīr 2B and Dhīr 5B can operate offline, facilitating document-based automation, knowledge retrieval, and enterprise services for sectors such as defense, government, industry, and education—making them particularly well-suited for areas with poor network connectivity or sensitive environments.

In a demonstration, ZenteiQ showcased test results for Dhīr 2B on a Motorola Edge 70 Fusion smartphone equipped with 12GB of RAM. Dhīr 5B, which was trained using Google’s TPU, was tested on the same smartphone—with the device set to airplane mode—using a 4-bit quantized model, thereby demonstrating its offline performance. Both models are available on AIKosha, an online repository for datasets and AI models launched in March 2025 by India’s Ministry of Electronics and Information Technology (MeitY).