The Global Model and Its India Problem
For the past few years, the conversation around artificial intelligence has been dominated by massive, powerful Large Language Models (LLMs) from global tech giants. While impressive, these models were trained primarily on English-language internet data,
reflecting Western culture and context. When applied to India, their limitations become apparent. The country's 22 official languages and countless dialects create a complex linguistic tapestry that global models struggle to navigate. Often, they resort to translating an Indian language query into English, processing it, and translating it back, a process that is inefficient, costly, and prone to losing nuance. This creates what is known as a 'token fertility' issue, where a single Hindi word might consume as many tokens as several English words, driving up costs significantly for Indian users.
The Rise of the Local AI Champions
In response to these challenges, a new ecosystem of Indian AI startups is emerging. Companies like Sarvam AI and Ola's Krutrim are not trying to out-muscle global giants on sheer size. Instead, they are building smaller, more efficient models designed from the ground up for India. Sarvam AI, for example, made waves with OpenHathi, one of the first open-source Hindi LLMs designed to be on par with models like GPT-3.5 for Indic languages. Similarly, Ola's Krutrim is developing a multilingual AI agent capable of understanding and responding in multiple Indian languages to perform real-world tasks. These companies are making a strategic bet: that a model deeply trained on Indian languages, dialects, and cultural contexts can deliver more value than a larger, more generic one.
What 'Adapted for India' Really Means
Adaptation goes far beyond simple translation. It means training models on vast amounts of high-quality data in languages like Hindi, Tamil, and Bengali. It involves designing systems that understand 'Hinglish' and other forms of code-switching, where speakers mix languages in a single conversation. It also means being cost-effective. Small Language Models (SLMs) are gaining traction because they are cheaper to run and can be deployed on more modest hardware, even smartphones, which is crucial for scaling in India's price-sensitive market. These models are not just technically different; they are built with a fundamentally different purpose—to be a practical, affordable tool for the Indian user, not a one-size-fits-all global product.
Unlocking Everyday AI Value
This shift towards localised AI is where the promise of 'everyday AI value' becomes a reality. The applications are tangible and transformative. Imagine a customer service chatbot for a bank that can seamlessly converse with a customer in rural Maharashtra in their specific Marathi dialect. Think of AI-powered healthcare apps that can help doctors in remote areas by analysing medical information in local languages, or an agricultural AI that provides pest alerts to farmers via voice notes in their native tongue. These aren't futuristic dreams; they are the specific, high-impact use cases that India-focused AI is being built to solve. This is AI that assists with government services, improves crop yields, and makes digital payments safer for everyone, regardless of the language they speak.
A Complement, Not Just a Competition
The rise of smaller, local AI does not mean the end of global models in India. Instead, the smartest path forward appears to be a hybrid one. Global LLMs can provide the broad reasoning power for complex problems, while smaller, specialised models handle the last-mile delivery, ensuring the technology is linguistically and culturally relevant. This approach allows India to leverage the best of both worlds: the immense scale and research of global players, and the precision, affordability, and cultural fluency of its own homegrown solutions. The goal is not necessarily to beat the global giants at their own game but to play a different one entirely—one focused on sovereign AI capabilities that serve the nation's unique needs.
