Decode Civils CAShorts
Current affairs Open app
Sci-TechPrelimsGS3

IndiaAI Mission releases indigenous foundation models EKA-1 and Dhir

By CAShorts DeskPublished Source: PIB (Ministry of Electronics and Information Technology, 3 Nov 2026; Union Cabinet, 7 Mar 2024; MeitY releases of 2025 and 2026)

IndiaAI Mission releases indigenous foundation models EKA-1 and Dhir
Illustration (AI-generated)

The IndiaAI Mission of the Ministry of Electronics and Information Technology (MeitY) released three indigenous foundation models on 3 November 2026 under its Innovation Centre pillar. EKA-1, the larger model, was developed by Soket Labs and has 26.3 billion parameters. Dhir 2B and Dhir 5B, developed by ZenteiQ AiTech, are open-source models that can run offline.

Syllabus: Prelims · Current Affairs & IR; GS3 · Science & Technology; GS3 · Indian Achievements in S&T

Why this matters for UPSC

GS3 science and technology: AI and the indigenisation of technology. Prelims can ask the IndiaAI Mission's pillars and model terms; Mains, India's case for sovereign AI.

Why it matters

A government office in a district with patchy internet could run a small model on its own computer and search its files in a local language, without sending data to a foreign cloud. That is the practical case for building models at home. Data stays inside Indian systems, languages that global models serve poorly get attention, and departments that handle sensitive records gain a tool they control. Smaller models also cost less to run because they need modest hardware. The larger question is whether public money for compute and research yields models good enough for wide use, since a model that is released but not adopted adds little to national capacity.

Prelims pointers

  • The IndiaAI Mission is implemented by IndiaAI, an Independent Business Division under the Digital India Corporation, under the Ministry of Electronics and Information Technology.
  • The mission's seven pillars are Compute Capacity, Innovation Centre, Datasets Platform (AIKosh), Application Development Initiative, FutureSkills, Startup Financing and Safe & Trusted AI.
  • The Innovation Centre builds indigenous models; the Datasets Platform, AIKosh, gives access to non-personal datasets and models; the Compute pillar provides shared GPU capacity.
  • A model's parameters are the weights it learns in training, while tokens are the pieces of text it is trained on; the two counts measure different things.
  • On-premise and on-device deployment keep a model and its data on the user's own hardware; cloud deployment sends each query to a remote data centre.

Practise a UPSC-style question

With reference to artificial intelligence language models, consider the following statements:

1. The parameters of a language model are the units of text, such as words or word pieces, on which it is trained.

2. A sparse model activates only a part of its parameters to process each input.

Which of the statements given above is/are correct?

  1. (a) 1 only
  2. (b) 2 only
  3. (c) Both 1 and 2
  4. (d) Neither 1 nor 2
Show the answer

Answer: (b) 2 only

Answer: (b) 2 only. Statement 1 describes tokens, not parameters.

Statement 1: incorrect. Parameters are the internal weights a model learns in training; tokens are the units of text it reads, as in EKA-1's 26.3 billion parameters and 7.5 trillion tokens.

Statement 2: correct. A sparse model routes each input through only part of its network, which saves computing power compared with a dense model of the same size.

Eliminate: statement 1 swaps tokens for parameters. Out: (a), (c).

Eliminate: statement 2 is a standard feature of sparse models. Out: (d).

Trap: Look-alike names. Parameters and tokens are both counted in billions or trillions, so a student who has only seen the numbers can mix up which word means what.

Remember: Parameters are learned weights; tokens are the text a model is trained on.

More from the 5 November 2026 edition

All stories of 5 November 2026

More Sci-Tech current affairs