Why India Is Betting on Smaller AI Models

News

1. Zuckerberg Is Developing an AI Agent to Assist Him as CEO: Report

2. CFOs Face Mounting Pressure to Balance AI Growth With Stability, Gartner Says

3. Delhi Assembly launches AI Chatbot Vidhan Sathi Ahead of Budget Session

4. OpenAI Acquires Astral to Expand AI Coding Capabilities

5. WhatsApp Plans Shift to Usernames from Phone Numbers

6. India Weighs Expanding Content Takedown Powers Across Ministries

India is betting on smaller, application-focused AI models as the fastest way to scale adoption across a fragmented, multilingual and infrastructure-constrained market, marking a clear break from the global race to build ever-larger systems.

That contrast is becoming more visible as global AI leaders continue to invest heavily in computing infrastructure, a trend evident at Nvidia’s GTC this week, where the company projected up to $1 trillion in AI chip demand and unveiled new systems to support large-scale AI workloads.

Small language models, or SLMs, are light, task-specific AI systems designed to run efficiently on limited computing resources, while large language models, or LLMs, are much larger systems trained on vast datasets to handle a wide range of tasks and typically require significant computing power and infrastructure.

That position has gained traction over the past year at forums such as the World Economic Forum in Davos in January and the India AI Impact Summit in New Delhi in February, where attention has shifted from building bigger models to making them work in real-world settings.

At the New Delhi summit, French President Emmanuel Macron praised India’s “deliberate” push toward efficient, sovereign AI designed for mass access via smartphones. At the same time, Indian Information Technology Minister Ashwini Vaishnaw argued in Davos that smaller models already handle most real-world AI workloads and that lean, application-focused systems, rather than massive architectures, will define the next phase of adoption.

India’s strategy centers on building affordable AI solutions aligned to specific use cases, Ashwini said in January.

That framing reflects a deeper structural reality. India is not attempting to outspend global technology firms in building ever-larger models, but to make AI usable across its linguistic, economic and infrastructural diversity.

Rishi Bal, Executive Vice-President and Head of BharatGen, a government-backed AI initiative focused on Indian-language models, said the distinction is not about scale for its own sake but about contextual relevance.

“In a global environment dominated by extremely large, general-purpose AI, the true value for India is not the sheer size of the model, but how well the AI fits India’s unique contexts,” he said.

 

Local Fit Matters

That gap between scale and usability is already visible in how global models perform in India.

Bal described LLMs as “architectural marvels,” but said they often operate as “cultural tourists” in the Indian context.

“Built on English-first tokenization (where text is broken into smaller units, or ‘tokens,’ that the model can process), these models frequently miss the socio-cultural nuances of our 22 scheduled languages and 19,500-plus mother tongues. They struggle to bridge the gap between a legal nuance in Marathi and a folk idiom in Maithili, inadvertently creating a new digital divide,” he said,  for enterprises, the issue is more practical

“India’s SLM strategy aligns perfectly with the future of enterprise AI, where outcomes triumph over bloated models,” said Umesh Sachdev, Co-Founder and CEO of Uniphore, a conversational AI and automation platform.

Dev Singh, Founder and CEO of Airo Digital Labs, an AI and data engineering firm, said India’s digital public infrastructure has created conditions that are difficult to replicate elsewhere.

“India is a unique blend of scale and diversity, and what works for the rest of the world may or may not work for India,” Singh said. “Its fast-growing digital public infrastructure has created the world’s largest AI consumer base in India.”

That combination of scale and constraint is precisely where smaller models gain an advantage.

Pavan Nanjudaiah, head of Tredence Studio, the AI and data science arm of analytics firm Tredence, said, “Domain-specific SLMs are significantly cheaper to train, fine-tune, and deploy, while delivering better performance on targeted use cases such as healthcare or governance in regional languages.”

They are also less prone to hallucinations in constrained environments and can operate on edge devices, such as smartphones or local machines, rather than relying entirely on remote cloud servers, making them viable in low-bandwidth settings.

“This is vital since a large share of villages still face variable connectivity,” he said.

Because these models are often developed and hosted within India, companies said they can be aligned more easily with local data protection requirements under the Digital Personal Data Protection Act, particularly when handling sensitive datasets such as identity, health and financial records. Their smaller, modular design also allows different specialized models to be combined for specific tasks instead of relying on a single large system.

 

Dual Track Strategy

India’s focus on smaller models does not signal a retreat from building large ones. Instead, it reflects a layered strategy.

Vaishnaw said the country is working on about 12 AI models, including larger ones in the 50 billion to 120 billion parameter range, aimed at handling more advanced enterprise tasks such as analytics, automation and decision support.

At the February summit, companies such as Sarvam AI, Gnani.ai and BharatGen rolled out a range of models for different use cases. These “sovereign” systems are developed and hosted within India, giving local control over data and deployment.

Sarvam AI introduced 30 billion and 105 billion parameter models aimed at multilingual voice and reasoning. Gnani.ai unveiled its VachanaTTS model designed for Indian languages and accents, capable of voice cloning from short audio samples, while BharatGen launched its 17 billion parameter Param2 model for specialized applications.

“India is done just consuming global LLMs; it’s building custom, sovereign, multilingual AI muscle,” Sachdev said.

Nanjudaiah described India’s SLM-plus-LLM approach as a “strategic inevitability,” combining large models for complex reasoning with smaller systems deployed at the edge for last-mile use cases.

He pointed to Sarvam AI’s 105 billion parameter model, which he said was able to match or outperform parts of China’s much larger DeepSeek model on certain tests while using fewer active components, suggesting efficiency gains rather than brute scale. However, such comparisons remain limited in public detail.

Bal said BharatGen follows a similar tiered architecture in which it uses a combination of large base models for general capabilities and smaller, specialized models for specific tasks.

Its foundational models provide broad multilingual capability, while domain-specific systems such as AyurParam, LegalParam, FinanceParam and AgriParam are deliberately smaller and trained on localized datasets for enterprise deployment on affordable, locally hosted hardware.

Jayaprakash Nair, global head of Data and AI Lab at Altimetrik, a digital engineering firm, said enterprises are increasingly moving toward such models.

“Starting with a capable base model, organizations fine-tune it using internal documentation and reinforce it with feedback from subject matter experts,” he said. “This enables accurate results within real business workflows and allows role-based customization, which is only economically viable with smaller models.”

Bal summarized the approach as a stack rather than a trade-off.

“It is a stack: large foundational models that provide deep reasoning and complex capability at scale, and small domain-specific models that deploy that capability where it matters,” he said.

Scale Versus Reach

The implications of that strategy are already visible.

Sunil Kansal, head of consulting and valuations at Shasat, said the approach lowers entry barriers, reduces reliance on expensive compute infrastructure and accelerates enterprise adoption across sectors such as banking, governance and education.

In the longer term, it could position India as a builder of sovereign, industry-focused AI systems rather than a consumer of global platforms.

Nanjudaiah said the model is expanding AI development beyond major cities, with labs operating in more than 27 Tier 2 and Tier 3 locations and enabling startups and state governments to build systems without high-end compute infrastructure.

He added that multilingual AI is beginning to reshape public service delivery, including voice-based Aadhaar interactions that allow citizens to access services in their mother tongue.

Over time, he said, India could export models designed for multilingual, mobile-first and low-resource environments across South Asia, Southeast Asia and Africa, adding that this shift would also build a talent base focused on creating AI systems from the ground up, rather than relying only on using existing tools developed by global firms.

However, he cautioned that the strategy carries risks if not balanced with investment in frontier research.

“While the US is building the most powerful AI, India is building the most accessible AI,” he said. “Both are valid forms of leadership, but the latter addresses a larger population.”

Without standardized benchmarks for Indian-language systems, he added, the ecosystem risks fragmentation and uneven quality even as adoption accelerates.