IndiaAI unveils EKA-1 and Dhīr, with access and evaluation details still to check

IndiaAI’s November 3 launch puts two different approaches to Indian-language AI in view: Soket’s larger EKA-1 and ZenteiQ’s compact Dhīr models. The announcement is official, but access and evaluation evidence need a closer look before teams plan a deployment.

The Ministry of Electronics and IT announcement identifies EKA-1, Dhīr 2B and Dhīr 5B as releases under the IndiaAI Innovation Centre. They were unveiled at a press event in New Delhi, supported by the mission’s compute and research programme.

EKA-1: a larger multilingual model, with a waitlist

According to the ministry, Soket’s EKA-1 has 26.3 billion parameters and was trained on approximately 7.5 trillion tokens. It is described as a sparse hybrid language model covering all 22 scheduled Indian languages and English, aimed at mathematics, coding, reasoning and agentic applications.

For prospective users, the clearest verified access route is Soket’s early-access waitlist. The page says the team reviews requests and emails applicants after approval. It also asks applicants to agree to the use of their prompts and responses to improve the models. Teams considering confidential material should examine those conditions before applying.

The launch announcement should therefore not be read as confirmation of unrestricted EKA-1 downloads. We have not verified a public weights repository or downloaded model files.

Dhīr: smaller models aimed at local use

ZenteiQ’s model page lists a 2.03-billion-parameter dense Dhīr 2B and a Dhīr 5B mixture-of-experts model with 5.09 billion total parameters, of which 1.79 billion are active per token. Both are text-only, with a 16,384-token context and support for English plus ten Indian languages.

The developer labels the release Apache 2.0 and links to AIKosh for downloads. We could not verify the files behind that destination during this review. Readers should confirm the actual checkpoint, accompanying licence and download requirements before treating it as deployment-ready access.

ZenteiQ positions Dhīr for offline and on-premises assistants. It also cautions that factual recall is a weakness, larger models lead on mathematics and code, and instruction following has been measured only in English.

Why there is no leading-model scorecard yet

ZenteiQ publishes developer-reported category averages and says it evaluated 41 benchmarks against 24 small open models. However, the linked technical-report page still says the report is forthcoming. The accessible summary does not provide the full peer-by-peer tables needed to verify a comparison against leading alternatives. These claims are not independent validation, and we have not converted them into a ranking.

A useful comparison needs exact model versions, individual test results, prompt settings and comparable inference conditions. Parameter counts and language coverage alone cannot establish which model will perform best on a particular task.

What to test before adoption

For a multilingual assistant, start with a small set of real questions in the languages it will serve, including mixed-language queries and cases where the correct response is to ask for clarification. Check whether answers remain grounded in supplied documents, rather than judging only fluent wording.

For local deployment, verify memory use, response time and reliability on the intended hardware. Keeping inference local can reduce data movement, but security still depends on the surrounding application, access controls and logging. A launch milestone is a reason to investigate these models, not a substitute for that evaluation.

Photo: Press Information Bureau, Government of India, November 3, 2026, photo 219468. Original launch-event photograph reproduced unchanged under PIB’s reproduction policy. Access and source pages checked November 5, 2026.