Quick answer
Economic Survey 2025-26 uses Chart XIV.7 to show open models narrowing the performance gap, and proposes AI-OS coordination for shared domestic innovation.
Open models permit some degree of inspection, adaptation or reuse, while small models reduce the resources needed for defined tasks; these are different characteristics. The Survey favours shared development and application-specific models to lower entry barriers, build domestic capabilities and reduce reliance on foreign proprietary systems.
Concept diagram
Different model routes, useful outcomes
Open and proprietary routes converge on task evaluation; size and openness are separate properties.
- Open weights do not automatically mean open data or zero deployment cost.
- Smaller task-specific models still require evaluation, maintenance and safeguards.
Basics: openness, size and adaptation are separate axes
Open-source versus open-weight
- Open-weight models release trained parameters subject to licence conditions; code, training data and full reproducibility need not all be available.
- Open-source claims should be checked against the actual licence and released components, not inferred from free access.
Small versus frontier
- Small models generally use fewer parameters and less compute than very large frontier models.
- Size does not determine whether a model is open or closed; narrow tasks can favour efficient specialised designs.
Fine-tuning versus inference
- Fine-tuning adapts an existing model through additional training; inference uses a trained model to produce outputs.
- Local inference may be practical on phones or PCs even when original model training required substantial shared infrastructure.
Chart XIV.7: what the evidence establishes
| Element | Survey reading | Limit on inference |
|---|---|---|
| Performance trend | Open models have been consistently closing the performance gap with closed models | Not proof that every open model equals every closed model |
| Frontier proximity | The best open models approach the frontier | Benchmark proximity is not universal safety or task suitability |
| Source and date | Ben Cottier et al. (2024), Epoch AI; report cited as published 24 November 2024 | Historical benchmark evidence, not a live 2026 league table |
| Numerical reproduction | No exact performance series transcribed here | The text extract does not expose reliable plotted point values |
Two complementary policy levers
- Openness spreads reusable knowledge and enables domestic developers to contribute improvements.
- Small, task-specific models can reduce hardware and operating constraints for deployment.
- Coordination tackles fragmentation in data quality, standards and interoperability rather than replacing distributed innovation.
AI-OS: proposed public coordination, not merely an operating system
Data foundations
Work with state and local institutions on structured, anonymised, machine-readable priority-sector datasets.
Shared compute
Pool existing data-centre capacity into common cloud infrastructure.
Collaborative repository
Create a government-hosted, community-curated platform under the IndiaAI Mission for sharing code and improvements.
Audit and scale
Coordinate and audit open efforts; attract large domestic firms able to absorb risk and scale useful applications.
Economic logic and implementation checks
Knowledge spillovers
- Shared code reduces duplication and helps entrants build on previous work; common standards improve compatibility.
- The Survey compares the sovereign’s monetary participation in AI-OS to UPI and Aadhaar, seeking public-good benefits.
Competition and accountability
- Domestic collaboration can lower dependence on proprietary foreign systems without implying technological autarky.
- A public repository still needs maintenance, licence compliance, security review and clear responsibility for failures.
Exam traps: distinguish claim from inference
Statements examiners like to twist — learn the correct version.
An open-weight model necessarily exposes its training data.
Weights, source code, training data and usage rights are separate dimensions of openness.
Chart XIV.7 proves open models dominate every task.
It supports narrowing performance gaps, not uniform superiority across benchmarks and deployments.
AI-OS is already a universally deployed government operating system.
The chapter proposes an initiative for coordinating a bottom-up AI ecosystem.
Prelims focus
- Chart XIV.7 cites Epoch AI’s 2024 open-model study, not an official Indian model-ranking exercise.
- AI-OS and the India-specific repository are proposed coordination mechanisms.
Mains angle
- Explain how open models and shared infrastructure can lower entry barriers while preserving incentives for private investment.
- Distinguish sovereign capability from complete self-sufficiency; discuss standards, auditability and sustainable maintenance.
Flip & recall
Answer in your head first, then tap the card to check.
Practice MCQs
Original Prelims-style questions with explanations.
0 / 3 correct · 3 left
Q1.Which statement correctly distinguishes model characteristics?
Q2.The principal inference from Chart XIV.7 is that:
Q3.Which is proposed as part of AI-OS?
