MeitY Launches Three New Sovereign AI Models Under the IndiaAI Mission

IndiaAI Mission was approved by the Union Cabinet in March 2024, under the aegis of the Ministry of Electronics and Information Technology (MeitY). The initiative was launched with the aim of “Making AI in India and Making AI Work for India”. Now, MeitY has launched three new indigenous AI models from two AI labs, namely Soket Labs Technology and Research and ZenteiQ AiTech Innovations. Soket AI claims that its AI model demonstrated “advanced capabilities” in math, coding, reasoning, and agentic use. Meanwhile, ZenteiQ says that its open-source large language models (LLMs) offer support for 10 Indian languages.

MeitY Releases EKA-1, Dhīr 2B and Dhīr 5B Sovereign AI Models

In a press release on Wednesday, MeitY announced that it has released three new indigenous AI models under the IndiaAI mission, the flagship initiative for AI innovation in India. The ministry launched EKA-1 from Soket AI and Dhīr 2B and Dhīr 5B from ZenteiQ. Soket claims that EKA-1 features 26.3 billion parameters.

On top of this, the new indigenous AI model is claimed to have been trained on around 7.5 trillion tokens. EKA-1 has been trained using 550 million Supervised Fine-Tuning (SFT) data samples across different domains. It features Soket's RALSNet (Recursive Attentive Latent Sparsity Network) hybrid architecture, which is claimed to deliver 3x higher throughput than models of similar size. The AI model also gets in-house pipelines for data crawling. OCR processing, labelling and filtering, multi-stage deduplication, tokenization, and data mixing.

Meanwhile, ZenteiQ's Dhīr 2B and Dhīr 5B are open-source AI models that can be deployed on edge devices as multilingual assistants. The AI models support English and 10 regional Indian languages. Both Dhīr 2B and Dhīr 5B can run offline for document-grounded automation, knowledge retrieval, and enterprise services spanning defence, government, industrial, and educational sectors in areas with poor connectivity and sensitive zones.

In its presentation, ZenteiQ showed that the Dhīr 2B was tested on the Motorola Edge 70 Fusion handset with 12GB of RAM. Trained on Google's TPUs, the Dhīr 5B was tested on the smartphone in flight mode with a 4-bit quantized model artifact, demonstrating its offline capabilities. Both models are available on AIKosha, an online repository for datasets and AI models, which was launched by MeitY in March 2025.