When Bigger Isn't Better
For the past few years, the artificial intelligence narrative has been about scale. Tech behemoths have been locked in a race to build Large Language Models (LLMs) with hundreds of billions, and even trillions, of parameters. The prevailing wisdom was
that bigger models, trained on more data, were inherently smarter. Yet, for a market as diverse and complex as India, this one-size-fits-all approach is proving to be inefficient and, at times, ineffective. Global LLMs, trained predominantly on English-language internet data, often struggle with the subcontinent's linguistic diversity and cultural nuances. They can be expensive to run, slow, and may not meet the data privacy and sovereignty requirements of enterprises in regulated sectors like banking and healthcare. This has created a significant opening for a new class of AI: smaller, nimbler, and hyper-focused.
The Small Model Advantage
Enter Small Language Models (SLMs). These are AI systems that are intentionally designed to be more compact. Instead of being a general encyclopedia on everything, an SLM is more like a specialized handbook for a specific job. They are trained on smaller, highly targeted datasets, which makes them faster, cheaper, and more efficient for specific enterprise tasks. According to a report from EY, SLMs are emerging as the practical foundation for AI growth in India. Their adaptability and efficiency make them ideal for a country with diverse languages and varying digital infrastructure. For businesses, the benefits are clear: faster performance, lower infrastructure costs, and the ability to customize models for domain-specific needs like financial regulations or medical diagnostics.
Lost in Translation, Found in Context
The core of the issue for global models is India’s linguistic reality. With 22 scheduled languages and thousands of dialects, a model trained primarily in English faces significant hurdles. It's not just about direct translation. It’s about understanding 'Hinglish'—the mix of Hindi and English common in urban conversation—regional accents, and cultural references that give communication its true meaning. Global models often fail at these tasks or are highly inefficient, generating more tokens and driving up costs for Indic languages. This is where India-focused models shine. By training on vast amounts of local language data, they can grasp context, sentiment, and intent far more accurately. This is crucial for applications ranging from customer service voice bots to delivering healthcare advice in a local dialect.
Building for Bharat
A new wave of Indian startups and tech giants is seizing this opportunity. Companies like Sarvam AI and Krutrim, both of which have achieved unicorn status, are leading the charge. Krutrim, launched by Ola's founder, is building a full-stack AI platform with models trained on extensive Indian-context data, supporting over 20 languages. Sarvam AI, on the other hand, is focused on enterprise solutions, building models optimized for voice applications and handling the nuances of Indian languages in real-time conversations. It has been selected under the IndiaAI Mission to help build the nation's sovereign LLM. Another key player is the government-backed Bhashini project, which aims to build a national public digital platform for languages, making AI accessible to all. These efforts are not about rejecting global AI, but about creating a hybrid path where large models might be used for broad reasoning while smaller, local models handle specialized, high-volume tasks.
The Road Ahead
The pivot towards smaller, context-aware AI models represents a strategic shift for India's tech ecosystem. It moves the country from being a mere consumer of global AI to an innovator building practical, compliant, and impactful solutions tailored to its own needs. For Indian enterprises, this means AI adoption becomes more affordable, secure, and effective. For a population where the vast majority do not speak English as their first language, it means technology that understands and responds to their lived realities. The future of AI in India may not be defined by who has the biggest model, but by who has the right model—one that truly speaks India's language.
