UPSC Prelims Practice Questions — Match AI models to workloads, not leaderboards
Q1. The IndiaAI Mission is operationalised on the ground by which one of the following entities?
- A. The IndiaAI Independent Business Division under Digital India Corporation
- B. The National e-Governance Division functioning as an autonomous society
- C. The Centre for Development of Advanced Computing, an autonomous R&D society
- D. The National Informatics Centre, the government's technology services arm
Q2. The national AI mission carrying an approved outlay of about ₹10,372 crore received its sanction from which authority, and in which year?
- A. The Cabinet Committee on Economic Affairs, in the year 2023
- B. The Union Cabinet, in the year 2024
- C. The Governing Council of NITI Aayog, in the year 2023
- D. The Cabinet Committee on Security, in the year 2024
Q3. Consider the following statements comparing open-weight AI models with closed, API-only models:
1. Open-weight models can be downloaded and executed on infrastructure controlled by the deploying organisation, whereas closed models are typically consumed through managed interfaces operated by the developing laboratory.
2. The release of a model's weights necessarily entails release of its entire training dataset and complete training source code under a free licence.
3. Self-hosting a model whose weights are available permits adaptation on proprietary data without that data leaving the organisation's own environment.
Which of the statements given above is/are correct?
- Open-weight models can be downloaded and executed on infrastructure controlled by the deploying organisation, whereas closed models are typically consumed through managed interfaces operated by the developing laboratory.
- The release of a model's weights necessarily entails release of its entire training dataset and complete training source code under a free licence.
- Self-hosting a model whose weights are available permits adaptation on proprietary data without that data leaving the organisation's own environment.
- A. 1 and 2 only
- B. 2 and 3 only
- C. 1 and 3 only
- D. 1, 2 and 3
Q4. With reference to the deployment of open-weight AI models by enterprises, consider the following statements:
1. The trained weights can be obtained and run on servers located within the deploying entity's own jurisdiction.
2. The model can be fine-tuned on an enterprise's confidential records without transmitting those records to the model's developer.
3. Availability of weights automatically confers unrestricted commercial rights on any user, since no licence conditions attach to such models.
4. On-premise hosting can be used to satisfy regulatory requirements that data be stored within national boundaries.
Which of the above is/are NOT correct?
- The trained weights can be obtained and run on servers located within the deploying entity's own jurisdiction.
- The model can be fine-tuned on an enterprise's confidential records without transmitting those records to the model's developer.
- Availability of weights automatically confers unrestricted commercial rights on any user, since no licence conditions attach to such models.
- On-premise hosting can be used to satisfy regulatory requirements that data be stored within national boundaries.
- A. 1 and 3
- B. 2 and 4
- C. 3 only
- D. 3 and 4
Q5. The India AI Governance Guidelines group their operative recommendations under how many pillars?
- A. Four
- B. Five
- C. Six
- D. Seven
Q6. The India AI Governance Guidelines, which recommend creating an AI Safety Institute and a Technology & Policy Expert Committee, were released by which one of the following?
- A. NITI Aayog, which had earlier issued the National Strategy for Artificial Intelligence
- B. The Department of Science and Technology, the nodal agency for national research missions
- C. The Ministry of Electronics and Information Technology, under the IndiaAI Mission
- D. The Telecom Regulatory Authority of India, the statutory regulator for digital communications
Q7. With reference to indigenous foundation models developed under the IndiaAI Mission, consider the following pairs of developer and description:
1. Sarvam AI — large language and speech models for Indian languages, with demonstrated accuracy in document understanding and Indic language processing.
2. Gnani AI — a 14-billion-parameter voice foundation model providing multilingual, real-time speech processing.
3. Soket AI — an open-source 120-billion-parameter foundation model targeting sectors such as defence, healthcare and education.
4. BharatGen — a consortium led by the Indian Institute of Technology, Madras.
Which of the above are correctly identified?
- Sarvam AI — large language and speech models for Indian languages, with demonstrated accuracy in document understanding and Indic language processing.
- Gnani AI — a 14-billion-parameter voice foundation model providing multilingual, real-time speech processing.
- Soket AI — an open-source 120-billion-parameter foundation model targeting sectors such as defence, healthcare and education.
- BharatGen — a consortium led by the Indian Institute of Technology, Madras.
- A. 1 and 3
- B. 2 and 4
- C. 1, 2 and 3
- D. 3 and 4
Q8. Consider the following statements:
1. Sarvam-30B is a Mixture-of-Experts model open-sourced under the Apache License, whereas the earlier Sarvam-1 was a 2-billion-parameter Indic model trained on an internal corpus of about 4 trillion tokens.
2. Sarvam AI is among twelve organisations selected under the IndiaAI Mission's foundation-model track, with financial and compute support of about ₹246.72 crore.
3. The open-source model built by Soket AI has a smaller parameter count than the voice foundation model built by Gnani AI.
Which of the statements given above is/are correct?
- Sarvam-30B is a Mixture-of-Experts model open-sourced under the Apache License, whereas the earlier Sarvam-1 was a 2-billion-parameter Indic model trained on an internal corpus of about 4 trillion tokens.
- Sarvam AI is among twelve organisations selected under the IndiaAI Mission's foundation-model track, with financial and compute support of about ₹246.72 crore.
- The open-source model built by Soket AI has a smaller parameter count than the voice foundation model built by Gnani AI.
- A. 1 and 2 only
- B. 2 and 3 only
- C. 1 and 3 only
- D. 1, 2 and 3
Q9. With reference to the India-AI Impact Summit 2026, consider the following statements:
1. It was hosted at Bharat Mandapam, New Delhi, and concluded on 20 February 2026.
2. Representation at the Summit was drawn from 118 countries.
3. It was preceded by about 550 pre-summit events held across some 30 countries.
4. Infrastructure-related investment pledges announced in connection with the Summit remained below 20 billion US dollars.
Which of the above is/are NOT correct?
- It was hosted at Bharat Mandapam, New Delhi, and concluded on 20 February 2026.
- Representation at the Summit was drawn from 118 countries.
- It was preceded by about 550 pre-summit events held across some 30 countries.
- Infrastructure-related investment pledges announced in connection with the Summit remained below 20 billion US dollars.
- A. 1 and 3
- B. 2 only
- C. 4 only
- D. 2 and 4
Q10. With reference to the design and outcomes of the India-AI Impact Summit 2026, consider the following statements:
1. Its main sessions were organised across seven thematic Chakras, one of which was 'Democratizing AI Resources'.
2. The Summit was anchored on three sutras, namely People, Planet and Progress.
3. Every participating country signed a legally binding treaty on AI safety, making compliance obligatory for all signatories.
4. Student engagement at the Summit crossed 2.5 lakh, recognised as a Guinness World Record.
Which of the statements given above are correct?
- Its main sessions were organised across seven thematic Chakras, one of which was 'Democratizing AI Resources'.
- The Summit was anchored on three sutras, namely People, Planet and Progress.
- Every participating country signed a legally binding treaty on AI safety, making compliance obligatory for all signatories.
- Student engagement at the Summit crossed 2.5 lakh, recognised as a Guinness World Record.
- A. 1, 2 and 3
- B. 1 and 4 only
- C. 1, 2 and 4
- D. 2, 3 and 4