India's IndiaAI Mission unveils three homegrown foundation models: EKA-1, Dhīr 2B and Dhīr 5B
MeitY says EKA-1 is a 26-billion-parameter model for maths, code and reasoning, while ZenteiQ's Dhīr models run offline on phones and ship with Apache 2.0 weights.
Why it matters
These are among the first releases from the IndiaAI Mission's foundation-model pillar, which gives domestic teams subsidised national compute. Officials stressed deployment in air-gapped and offline settings for government and enterprise rather than competing with the largest frontier models. EKA-1 was trained on 7.5 trillion tokens and covers more than 100 languages; the Dhīr models were trained from scratch on Google Cloud TPUs and support English and 10 Indian languages. They follow Sarvam's Prism release earlier on Tuesday. Watch for independent benchmarks once the weights are on AIKosh.
Key facts
- EKA-1: 26 billion parameters, trained on 7.5 trillion tokens
- EKA-1 supports more than 100 languages
- Dhīr 2B and 5B are mixture-of-experts models trained on Google Cloud TPUs
- Dhīr supports English and 10 Indian languages and runs offline on phones
- Dhīr weights released under Apache 2.0 via AIKosh
- Announced by MeitY Secretary S. Krishnan on 3 November