India releases EKA-1 and Dhīr AI models

The released models are EKA-1, Dhīr 2B and Dhīr 5B.[1] Soket Labs Technology and Research developed EKA-1, while ZenteiQ AiTech Innovations developed the two Dhīr models.[1] Their development was enabled by subsidized access to national AI compute and support for ancillary research expenses.[1] MeitY Secretary S. Krishnan and senior ministry and IndiaAI officials attended the New Delhi launch.[1] Krishnan said indigenous models strengthen India’s technological and strategic autonomy in AI.[1]

Soket says EKA-1 has 26.3 billion parameters and was trained on approximately 7.5 trillion tokens.[1] It supports all 22 scheduled Indian languages as well as English, and is described as an open, sparse hybrid language model designed for mathematics, coding, reasoning and agentic use.[1] Soket says it was tailored to India’s banking, financial services and cybersecurity landscape.[1] The Dhīr models support English and 10 Indian languages; ZenteiQ says they support document-grounded automation, knowledge retrieval and citizen or enterprise services.[1] Dhīr 5B has 5.09 billion total parameters, with 1.79 billion active per token, according to ZenteiQ.[3]

ZenteiQ says the Dhīr models can be deployed on-premises, work without continuous cloud connectivity and use a lightweight architecture that allows deployment on mobile phones.[2] The company says local use keeps user data on devices, which it says enhances privacy, security and accessibility.[2] ZenteiQ lists both models under the Apache 2.0 license and says they are available to download through AIKosh; enterprise access is offered through licensing and pilot programmes.[3]

ZenteiQ says Dhīr 2B ranked first in its class on 13 of 15 Indic benchmarks, and says both Dhīr models were compared with 24 open models containing 1–7 billion parameters.[3] The company identifies fact recall without a supporting source as the models’ weakest area and recommends authoritative retrieval and human review in fields such as finance.[3] ZenteiQ also says neither Dhīr model has a reasoning mode.[3] IndiaAI COO Sudeep Shrivastava characterized the launch as progress spanning lightweight models for everyday devices and models for complex enterprise applications.[1]