Why in News?
On 3 November 2026, the IndiaAI Mission released Soket’s EKA-1 and ZenteiQ’s Dhīr models under its Innovation Centre pillar.
- The ministry reports that EKA-1 covers all 22 Scheduled Indian languages and English. Soket designed the model for tasks including mathematics, coding and reasoning.
- ZenteiQ introduced Dhīr 2B and Dhīr 5B as multilingual assistants supporting English and ten Indian languages, with the option to run them on local equipment.
- Both teams received access to subsidised computing resources and support for research expenses under the mission.
- An agency choosing an assistant needs to establish which languages its users speak and where their requests will be processed. These are separate procurement decisions.
- The announcement describes developer capabilities. Evaluating a particular public service requires testing the model on the documents, language and hardware used there.
UPSC Relevance
Prelims Relevance
- The IndiaAI Innovation Centre supports indigenous foundation models.
- A token is a unit into which an input is divided for processing.
- Training adjusts model parameters using examples.
- Inference applies a trained model to a new input.
Mains Relevance
GS Paper 3
- Examine how shared computing infrastructure can lower barriers to indigenous model development.
- Explain why language coverage and deployment conditions require separate evaluation.
Background and Context
What Does Multilingual Training Establish?
A foundation model learns patterns across a broad collection of examples and can be adapted or prompted for several tasks.
- During training, a model processes sequences of tokens and adjusts internal parameters. A token can represent part of a word, so a token count cannot be read as a count of words or documents.
- The ministry reports 26.3 billion parameters for EKA-1. These are learned numerical values. The figure describes model size, while a measure of accuracy needs a task, test examples and a scoring rule.
- A language coverage claim tells a department which languages to consider for a trial. It leaves open how well the model handles a local spelling, an unfamiliar abbreviation or a question mixing two languages.
- Consider an eligibility question asked in two languages. A useful paired test keeps the facts and qualifying conditions the same, then checks whether each answer preserves the exception that determines the applicant’s eligibility.
How Can a Local Assistant Answer Offline?
Once a trained model and its supporting software are installed, a local machine can process new requests using its own computing resources.
- During local inference, a device or institution’s server produces the response. A field office can process requests during an internet outage if the application and its reference material are available locally.
- With document grounding, an application retrieves relevant passages and supplies them to the model alongside the question. Staff can compare the response against the passage to see whether it supports the answer given.
- A remote document index can become unreachable during an outage. An agency should test the whole application, including how it stores and retrieves the circular a user asks about.
- Keeping processing on the premises can limit where requests travel. Access controls still matter because a shared workstation, an exposed local index or retained query histories can reveal sensitive information to another user.
What Should a Department Test Before Adoption?
A department can test both a language model and its supporting application using questions that staff encounter in daily work.
- For a multilingual enquiry service, build a reference set with fluent speakers and subject specialists. Include an ordinary request and a borderline case in which a small translation error changes who qualifies.
- For a field office, measure response time and memory use on the equipment staff will have. Repeat the trial with network access disabled and record which functions remain available throughout an ordinary request.
- Ask the assistant about an amended circular after replacing the old copy. Check whether retrieval finds the current wording and whether the answer preserves its effective date and any exceptions.
- Route uncertain answers to a responsible officer. A model may generate fluent language even when the relevant record is absent, and an applicant needs a way to obtain an authoritative decision.
Way Forward
Set Acceptance Criteria for Each Service
- Publish language-specific test results for the tasks that matter to users.
- Assign responsibility for refreshing the local document collection when instructions change.
- Test access restrictions and an officer referral before using the assistant with sensitive enquiries.
Conclusion
- The IndiaAI release gives agencies two practical questions to investigate. Which model handles the language of the intended user, and can the complete application run where that user needs it?
- A defensible choice rests on task-level evidence. An agency can demonstrate whether an assistant preserves a condition in translation, retrieves the current rule and remains usable through a network outage.
UPSC Practice Questions
Prelims MCQ 1
With reference to the IndiaAI model release, consider the following statements:
- EKA-1 covers all 22 Scheduled Indian languages according to the ministry.
- The Dhīr assistants are described as capable of local deployment.
- Language coverage establishes equal accuracy across every supported language.
How many of the above statements are correct?
(a) Only one (b) Only two (c) All three (d) None
Answer: (b) Only two
Explanation:
Statements 1 and 2 reflect the announced capabilities. Equal accuracy would require comparable task-level testing in each language.
Prelims MCQ 2
A district office runs a trained model on its own server to answer a new question. Which operation is it performing?
(a) Pretraining (b) Updating model parameters (c) Local inference (d) Labelling a training corpus
Answer: (c) Local inference
Explanation:
The office uses the existing model to produce a response on local equipment.
UPSC Mains Questions
- How can India assess whether indigenous multilingual foundation models improve access to public services?
- Examine the operational and governance conditions for deploying an offline language assistant in district administration.
Source: PIB, Ministry of Electronics and Information Technology.
Frequently Asked Questions
Who developed the models released by IndiaAI?
Soket developed EKA-1. ZenteiQ developed the Dhīr 2B and 5B assistants. Both teams received support under the IndiaAI Innovation Centre pillar for developing indigenous foundation models.
How does language coverage differ between the models?
The ministry reports coverage of all 22 Scheduled Indian languages and English for EKA-1. It describes the Dhīr assistants as supporting English and ten Indian languages.
What does local inference require?
The device or local server needs the trained model, suitable computing resources and the software to process requests. Offline document lookup also needs a local copy of the required reference material.
How can officials test answers grounded in documents?
Officials can select an enquiry with a known answer, inspect the retrieved passage and compare the response with the current circular, including its exceptions and effective date.
Tell Google you want more of this.
Add Anantam IAS as a preferred sourceOne tap, and this site shows up more often in your own Top Stories, AI Overviews and AI Mode. Remove it any time.
