The ministry says EKA-1 has 26.3 billion parameters and was trained on roughly 7.5 trillion tokens, the text units a model processes. It targets maths, coding and reasoning across all 22 constitutionally scheduled Indian languages and English.
Dhīr models are announced as open source, supporting English and ten Indian languages, with local execution without an internet connection. Developers target sensitive on-premise applications among other uses. The consulted sources provide no independent comparison establishing their performance.