TARGET FOR IQ · CURRENT AFFAIRS
IndiaAI Models EKA-1 and Dhīr: November 2026 Release and Key Differences
Current affairs reading · 3 min read · English
On this page
What is the update?
IndiaAI announced EKA-1 and the Dhīr 2B and 5B models on 3 November 2026 under its foundation-model development effort. The releases address different language and deployment requirements within the IndiaAI Innovation Centre pillar.
The model comparison
| Model | Developer | Features reported at release |
|---|---|---|
| EKA-1 | Soket Labs | 26.3 billion parameters; about 7.5 trillion training tokens; 22 scheduled Indian languages plus English; sparse hybrid design. |
| Dhīr 2B and 5B | ZenteiQ | English and ten Indian languages; smaller models aimed at local, offline-capable and document-grounded applications. |
The announcement described EKA-1 use cases in reasoning, code and agentic tasks, while Dhīr was positioned for on-premise assistants in sectors including government, defence, industry and education. These are release descriptions, not independent comparative test results.
Parameters, tokens and deployment
Parameters are the learned values within a model. Training tokens describe units of data processed during training. The two numbers measure different things and should not be compared as though they were the same quantity.
A smaller model can have practical advantages where memory, processing power or connectivity is limited. A larger model may address different requirements. Model size alone, however, does not establish accuracy, suitability or the quality of answers in a particular language.
What local operation changes
On-premise or offline-capable deployment can allow an organisation to run an application within its own environment. This can affect data flows and access arrangements, but it does not automatically make every deployment secure or every answer reliable. Configuration, permissions, evaluation and ongoing maintenance still matter.
Document-grounded retrieval can help an application use a defined collection of material. It does not remove the need to check whether the retrieved passage is relevant or whether the generated answer represents it correctly.
The mission context
The government said subsidised compute and research support helped development. IndiaAI’s seven pillars cover compute, innovation, datasets through AIKosh, applications, future skills, startup financing and safe and trusted AI.
For exams, connect the release with public research infrastructure, Indian-language technology and strategic capacity. Distinguish the availability of a model from proof of performance across every intended use. A responsible comparison would specify the task, language, test set and deployment conditions.
Exam revision
- EKA-1: Soket; Dhīr: ZenteiQ.
- The models were announced under the IndiaAI Innovation Centre pillar.
- Parameter count and training-token count are different metrics.
Quick questions
Does local deployment automatically guarantee privacy and accuracy?
No. It changes where the system runs, but security controls and task-specific evaluation are still necessary.
Official references: Ministry of Electronics & IT — 2026-11-03.