News in short
- MeitY has released three indigenous AI models under the IndiaAI Mission.
- Soket’s EKA-1 is claimed to have 26.3 billion parameters and 7.5 trillion training tokens.
- EKA-1 supports 22 scheduled Indian languages, according to the announcement.
- ZenteiQ’s Dhīr 2B and Dhīr 5B are open-source models supporting 10 Indian languages.
- The Dhīr models can run offline on edge devices and are available through AIKosha.
The Ministry of Electronics and Information Technology (MeitY) has released three indigenous AI models under the IndiaAI Mission. The models come from Soket Labs Technology and Research and ZenteiQ AiTech Innovations.
The release includes Soket’s EKA-1 and ZenteiQ’s Dhīr 2B and Dhīr 5B. The models target different use cases, from reasoning and coding to multilingual assistance that can operate without an internet connection.
The IndiaAI Mission was approved by the Union Cabinet in March 2024. Its stated goal is to make AI in India and make AI work for India.
Soket EKA-1 claims 7.5 trillion training tokens
Soket says its EKA-1 model has 26.3 billion parameters and was trained on around 7.5 trillion tokens. The company also says it used 550 million supervised fine-tuning (SFT) data samples across multiple domains.
EKA-1 uses Soket’s RALSNet, or Recursive Attentive Latent Sparsity Network, hybrid architecture. The company claims the architecture delivers three times higher throughput than models of a similar size.
Soket says EKA-1 demonstrated advanced capabilities in mathematics, coding, reasoning and agentic use. The model also relies on in-house data pipelines for crawling, optical character recognition (OCR), labelling, filtering, deduplication, tokenisation and data mixing.
The company says EKA-1 supports 22 scheduled Indian languages. These are company-reported capabilities; the announcement does not provide independent benchmark results.
ZenteiQ Dhīr models bring offline AI to edge devices
ZenteiQ’s Dhīr 2B and Dhīr 5B are open-source large language models designed for deployment on edge devices. They support English and 10 regional Indian languages.
The models can operate offline for document-grounded automation, knowledge retrieval and enterprise services. ZenteiQ identifies potential use cases across defence, government, industry and education, particularly in areas with limited connectivity or sensitive operating conditions.
During its presentation, ZenteiQ showed Dhīr 2B running on a Motorola Edge 70 Fusion smartphone with 12GB of RAM. The company also demonstrated Dhīr 5B on the same phone in flight mode, using a 4-bit quantised model artifact.
Both models are available on AIKosha, MeitY’s online repository for datasets and AI models. The repository launched in March 2025.
The release adds to India’s effort to develop AI systems locally. The practical value of these models will depend on their performance across languages, devices and real-world tasks, as well as how developers adopt them.
