- Why in news
- On 3 November 2026, the IndiaAI Mission released the indigenous foundation models EKA-1 (Soket) and Dhīr 2B and 5B (ZenteiQ) in New Delhi. Both teams were chosen under the IndiaAI Innovation Centre pillar.
- What it is
- The IndiaAI Mission is a Central programme under MeitY that builds an AI ecosystem in India. Its Innovation Centre pillar backs teams that build indigenous foundation models and Large Language Models. A foundation model is a large AI model trained on broad data, which can then be adapted to many tasks.
- Mission approval
- The Union Cabinet approved the IndiaAI Mission on 7 March 2024.
- Seven pillars
- The mission has seven pillars: Compute Capacity, Innovation Centre, Datasets Platform (AIKosh), Application Development Initiative, FutureSkills, Startup Financing, and Safe & Trusted AI.
- Dhīr 2B and 5B (ZenteiQ)
- These are open-source models for secure, on-premise use. They support English and 10 Indian languages.
- Compute support
- Subsidised national AI compute and support for ancillary research costs enabled the models.
- AIKosh
- AIKosh is the IndiaAI datasets platform, and the models can be made available through it.
- Outlay and approval
- The Mission was approved in March 2024 with an outlay of ₹10,371 crore over five years.
- Responsible AI projects
- 13 Responsible AI projects were approved in educational institutions, covering bias mitigation, machine unlearning, privacy-preserving AI, explainability, deepfake detection and risk assessment.
- AI Safety Institute
- It was announced in January 2025 and works through a hub-and-spoke model; it oversees projects selected under the Expressions of Interest rounds.
- Governance Guidelines
- They set a risk-based framework and propose AIGEG, TPEC and AISI as institutional mechanisms.
- Summit deliverables
- India chaired the India AI Impact Summit 2026 in New Delhi; outcomes included Frontier AI Impact Commitments, a Guidance Note on AI Governance and the Trusted AI Commons.
Don’t confuse with EKA-1 vs Dhīr: EKA-1 is a large 26.3-billion-parameter model covering all 22 scheduled languages. Dhīr 2B and 5B are lightweight models for on-device use, supporting 10 Indian languages.
Don’t confuse with Foundation model vs LLM: A foundation model is the broader class of large pre-trained models. An LLM is a foundation model focused on language.
Don’t confuse with AIKosh vs IndiaAI Compute: AIKosh is the datasets and models platform; IndiaAI Compute provides GPU access.
Don’t confuse with AI Safety Institute vs AI Impact Summit: The Institute is a standing national body; the Summit was a one-time global event India chaired in 2026.
Science and technology: 57 Prelims questions since 2013; most test the concept behind the news, not the news itself. In Mains, 12 news-triggered questions.
Why it matters
- Home-built models strengthen technological and strategic autonomy in AI.
- Open, on-premise and on-device models suit sensitive sectors such as defence and government, and low-connectivity areas.
- Builds indigenous tools and standards for bias, privacy and deepfakes, raising public trust in AI.
- Gives India a voice for Global South priorities in AI governance.
Problems
- Indian-language training data is limited and uneven, and models need continued access to expensive compute.
- Guidelines are risk-based and largely voluntary, so enforcement is weak.
- Deepfakes and bias evolve faster than academic projects and standards.
Way forward
- Expand access to compute and datasets through AIKosh, and encourage adoption in public services and industry.
- Back guidelines with clear legal duties, audits and grievance redress.
- Scale tested tools from labs to deployment, with regular risk assessment.
Put in your answer
IndiaAI Mission approval: The Cabinet approved the mission on 7 March 2024, with seven pillars covering compute, models, data, skills, startups and safety.
Mission outlay: ₹10,371 crore over five years, approved March 2024.
Safe & Trusted AI projects: 13 Responsible AI projects cover bias, privacy, explainability and deepfake detection.
- 2026Mains GS3 · Q16Explain agentic AI, its working, applications, advantages, risks and challenges.
- 2026Mains GS3 · Q9Discuss fake news threats and 2021 IT Rules amendments on digital media.
- 2019Mains ESSAY · Q8Essay on whether AI brings jobless future or better jobs via reskilling.
- 2026Mains GS4 · Q1Discuss ethics and accountability of professor using AI to generate evaluation report.
- 2026Mains GS4 · Q8Assess AI predictive policing using biometric data of crowds, ethically and legally.
- 2020Prelims Q38Which AI capabilities are achievable at present state of development?
UPSC’s own questions that come close to this issue, from our Prelims and Mains banks.
Prelims practice · Statements
Ours, not from a UPSC paper
Consider the following statements:
- 1.The IndiaAI Mission was approved in March 2024 with an outlay of about ₹10,371 crore over five years.
- 2.The IndiaAI Mission has seven pillars, and Safe & Trusted AI is not one of them.
- 3.The IndiaAI Safety Institute was announced in January 2025 and engages stakeholders through a hub-and-spoke model.
Which of the statements given above is/are correct?
True or false?
1 of 7
Statements the way Prelims prints them. Say first, then see the fact.
1.The IndiaAI Mission was approved in March 2024 with an outlay of about ₹10,371 crore over five years.
True. Approved March 2024; ₹10,371 crore over five years.
2.The IndiaAI Mission has five pillars, and Safe & Trusted AI is not one of them.
False. It has seven pillars, and Safe & Trusted AI is one.
3.Under the Safe & Trusted AI pillar, projects have been approved on bias mitigation, machine unlearning and deepfake detection.
True. The 13 Responsible AI projects include these topics.
4.The IndiaAI Safety Institute was announced in January 2025 and follows a hub-and-spoke model.
True. Announced January 2025; hub-and-spoke engagement with stakeholders.
5.The India AI Governance Guidelines propose a single binding AI statute with criminal penalties for all AI misuse.
False. They set a risk-based framework with institutional mechanisms (AIGEG, TPEC, AISI), not a single statute.
6.Machine unlearning refers to techniques that remove unwanted or sensitive information from trained AI models.
True. Machine unlearning removes specific data or concepts from a model.
7.Federated learning trains models by collecting all raw user data in one central server.
False. Federated learning trains models across devices or sites without centralising raw data.
Written by us from the release, not by UPSC; checked by a second model.
Mains practice
15 marks · 250 words
The Safe & Trusted AI pillar of the IndiaAI Mission funds work on bias, privacy, explainability and deepfakes, and the India AI Governance Guidelines propose a risk-based framework instead of a single AI law. Examine the main provisions of the IndiaAI Mission and critically analyse the challenges India faces in ensuring safe and trusted AI.
Worth revising until Prelims 2028: most news questions come within two years of the event.