India's Ministry of Electronics and Information Technology announced on November 3, 2026 that the IndiaAI Mission had released three foundation models built by companies in its Innovation Centre pillar: EKA-1 from Soket Labs, and Dhīr 2B and 5B from ZenteiQ. The ministry says subsidised national compute and support for research costs enabled the work. MeitY Secretary S. Krishnan said such models strengthen "India's technological and strategic autonomy in Artificial Intelligence".
What was released
- EKA-1. The ministry describes an open, sparse hybrid model of 26.3 billion parameters, trained on about 7.5 trillion tokens, covering all 22 scheduled Indian languages and English, for mathematics, code, reasoning and agents. Its AIKosh card adds 52 layers combining attention, recurrent and expert layers, a 64K-token context window and a 30.49 GB download.
- Dhīr 5B. AIKosh lists a 5.09-billion-parameter Mixture-of-Experts model, with about 1.79 billion parameters active per token, supporting English and ten Indian languages, at 9.50 GB. ZenteiQ pitches both Dhīr models for on-premise and offline use. Its CEO says their lightweight design allows on-device deployment on phones.
Soket's CEO described EKA-1 as "specifically tailored to India's unique banking, financial services and cybersecurity landscape", which is a narrower pitch than the ministry's general description. The AIKosh card files it under the sector "Defence and National Security".
Reading the listing as well as the press release
The announcement says these models "can be made accessible" through AIKosh, and the listings show why that wording deserves care.
- License. Both the EKA-1 and Dhīr 5B cards say Apache 2.0, which is a genuinely permissive grant if the weights can be obtained.
- Access. Both cards also show "Visibility Restricted", without explaining what that means for a downloader. EKA-1's card asks developers to "apply for early access", and when we checked on November 9 it showed 0 downloads, against 38 for Dhīr 5B. We did not attempt a download.
- Evidence. EKA-1's card says "Internal evaluations indicate performance on par with comparably sized models, with detailed methodology and results to follow in a technical report." Inc42 noted that the announcement gave no benchmark results, deployment details or pricing.
- Small inconsistencies. The ministry says 26.3 billion parameters and 22 languages plus English, while the card says 26 billion and more than 100 Indian, Global South and programming languages. Soket's claim of more than three times faster inference is its own.
What remains uncertain. We cannot say how well these models perform in any Indian language, since no independent results exist yet. Neither the training data's composition nor the compute actually used was published in the materials we read. The restricted status may be a temporary platform setting or a real access limit.
Analysis: A domestically trained 26-billion-parameter model is a real engineering milestone, and it should now be tested the way any other model would be. If Geetha's team at Belgavi's Book Store wanted Kannada and Hindi support for return requests, a small labelled set of real customer questions would be more informative than any launch claim. Check the license on the weights you actually receive, measure language by language rather than in aggregate, and treat the technical report, when it arrives, as the start of evaluation.
Sources and evidence
IndiaAI released Soket's EKA-1 and ZenteiQ's Dhir models on November 3, with Apache 2.0 licenses shown on AIKosh. But the cards are marked restricted, EKA-1 asks developers to apply for early access, and its evaluation is promised later. Treat this as a compute and capacity milestone, and test each language on your own data before relying on it.