Tech Explained: Why India’s AI strategy is betting on smaller models in Simple Terms

A new white paper released by the Office of the Principal Scientific Adviser to the Government of India suggests the country is prioritising smaller, specialised AI systems designed for real-world deployment, even as large frontier models continue to dominate global headlines.

The paper, titled “Advancing Indigenous Foundation Models”, was released on March 13 as part of the government’s AI Policy White Paper Series aimed at shaping India’s artificial intelligence ecosystem.

The document identifies the development of indigenous foundation models as a strategic priority to support inclusive growth and public good while aligning AI systems with India’s legal framework, societal values and national security interests.
Foundation models are large AI systems trained on vast datasets such as text, images, audio and video, enabling tasks like translation, summarisation, question answering and text classification.

The white paper says that building such models using India-relevant datasets and governance frameworks is critical to ensuring transparency, inclusivity and national alignment in the country’s AI ecosystem.

A shift toward smaller, sector-specific AI

While large language models remain important, the white paper highlights the growing importance of Small Language Models (SLMs) as a practical path for India.

These models are designed to be specialised and deployment-efficient, making them better suited for sectors such as agriculture, healthcare, education and micro, small and medium enterprises (MSMEs).

Compared with massive frontier models, smaller systems require far less compute and energy, making them significantly cheaper to deploy in real-world applications.

Not to forget, the paper does not frame this as a strategic shift away from large models. Instead, it presents SLMs and large foundation models as complementary layers within the same ecosystem rather than replacements for one another.

Also read: Sarvam AI launches indigenous 30B and 105B LLMs to power India’s sovereign deployment

The strategy also includes multimodal AI systems capable of processing text, speech and images, enabling applications in areas such as climate monitoring, public service delivery and urban governance.

Several indigenous initiatives are already emerging. Sarvam AI has introduced Sarvam-105B, a language model optimised for Indic languages, while Gnani.ai has launched Inya VoiceOS, a voice-to-voice AI system capable of processing speech directly across multiple languages.

Academic programmes are also contributing to the ecosystem. The BharatGen initiative led by the Indian Institute of Technology Bombay has developed models such as Param-1 for text processing, Shrutam for speech recognition, Sooktam for text-to-speech and Patram for document understanding.

Private companies are building smaller models for enterprise use as well. Zoho has introduced its in-house Zia LLM for enterprise workflows such as data extraction and summarisation, while CoRover.ai has launched BharatGPT, a multilingual conversational model trained on Indian conversational datasets.

Compute, data and language layers powering India’s AI ecosystem

To support indigenous model development, the government is building a multi-layered AI ecosystem combining compute infrastructure, datasets, language technologies and domestic foundation models.

At the compute layer, India is building shared AI infrastructure through the IndiaAI Mission, which was approved by the Union Cabinet in March 2024 with an outlay of ₹10,371.92 crore over five years and is being led by the Ministry of Electronics and Information Technology. The programme aims to strengthen domestic AI capabilities through shared compute and data infrastructure.

Under this initiative, the IndiaAI Compute Portal is providing large-scale GPU access for AI development. More than 38,000 GPUs have already been onboarded, with subsidised access priced at around ₹65 per hour, enabling startups, researchers, students and government agencies to train models at significantly lower cost.

Also read: India wants 200,000 GPUs in the government’s AI arsenal

Alongside compute infrastructure, the ecosystem includes AIKosh, a national repository designed to host datasets, models and other AI resources required for training foundation models.

At the language layer, the Bhashini initiative is building benchmarks and evaluation frameworks for speech and language AI across India’s linguistic diversity. Together, these efforts aim to build datasets, benchmarks and models covering all 22 scheduled Indian languages and dozens of dialects, ensuring that AI systems reflect India’s linguistic and cultural diversity.

The model layer includes emerging indigenous systems such as BharatGen, Sarvam and other domestic foundation-model initiatives, creating a pipeline from datasets and benchmarks to deployable AI systems.

The IndiaAI Mission has also launched structured innovation programmes to accelerate model development. A Call for Proposals issued in January 2025 invited startups, researchers and entrepreneurs to build foundation models tailored to India’s needs, receiving 506 proposals by April 2025.

In the first phase, four organisations, including Sarvam AI, Soket AI, Gnani.ai and Gan AI, were selected to develop multilingual text models, voice AI systems and text-to-speech technologies.

A second phase announced in September 2025 approved eight additional projects focused on building both large and small language models trained on Indian datasets covering all 22 scheduled languages.  The eight organisations selected under the second phase include Tech Mahindra, Fractal Analytics, Avataar.ai, Zenteiq AI, Genloop Intelligence, NeuroDX, Shodh AI, and the BharatGen consortium led by Indian Institute of Technology Bombay.

The white paper also explores a possible framework for AI training data governance. It discusses the idea of a “blanket licence” system where AI developers can train models on lawfully accessed content while paying royalties if the systems are later commercialised.

ALSO READ | India’s new AI rules giving platforms just three hours to remove flagged content raise fears of over-censorship

Such an approach could balance large-scale AI development with compensation for creators and rights holders.