The Limits of a Global Approach
For much of the recent AI boom, the focus has been on creating massive Large Language Models (LLMs) trained on vast swathes of the internet. While powerful, these global models often stumble in the uniquely complex Indian market. Their primary training
data is overwhelmingly English, making them less effective when dealing with India's 22 officially recognised languages and thousands of dialects. When they do process Indian languages, they often rely on inefficient translation, which increases costs and fails to capture nuance. This is a critical failure in a country where only a fraction of the population is fluent in English, and where code-switching—mixing languages like Hindi and English in a single sentence—is commonplace. A model that doesn't understand this reality cannot be practically useful for the majority of Indians.
Why Language Is More Than Words
The challenge goes beyond simple translation. True understanding requires cultural context. An AI assistant must know that in India, 'Thalaiva' could refer to the actor Rajinikanth or the cricketer MS Dhoni, depending on the user's location and context. Global models, lacking this deep-seated cultural fluency, can provide responses that are technically correct but practically useless or even nonsensical. Furthermore, much of India's cultural and traditional knowledge has been passed down orally and has never been extensively digitised, creating a massive gap in the training data available to global LLMs. This scarcity of high-quality, structured data in Indian languages means that models trained on global internet data are inherently biased and incomplete.
The Rise of the Niche and Nimble
This is where smaller, India-focused AI models are building a powerful case. Startups like Sarvam AI and Krutrim, along with consortiums like BharatGen, are developing models trained specifically on Indian languages and data. These models are not trying to be everything to everyone. Instead, they are being optimised for specific domains and languages. For example, Sarvam AI focuses on enterprise solutions and highly accurate voice-to-text for regional dialects, while Krutrim targets the consumer space with broad multilingual support. These Small Language Models (SLMs) offer clear advantages: they are faster, more cost-efficient to run, and can be fine-tuned for specific business needs, such as a banking chatbot that understands financial slang in Marathi or a healthcare app that can process patient queries in Bengali.
A Stronger Business Case
For enterprises, the shift from experimentation to production is making these practical benefits undeniable. As businesses integrate AI into their core operations, factors like cost, speed, and data privacy become paramount. SLMs have lower inference costs, meaning the day-to-day expense of running them is significantly less than for a massive global model. Their smaller size also allows them to be deployed on private servers or even edge devices, a crucial feature for regulated sectors like finance and healthcare that require data sovereignty and compliance. The ability to provide a genuinely localised customer service experience, generate marketing content that resonates with local festivals and trends, or create personalised user journeys in a customer's native language is a powerful driver of business value.
