For much of the generative
artificial intelligence (AI) boom, the conversation centred on building ever-larger language models. Companies raced to develop frontier models with hundreds of billions of parameters, while enterprises experimented with tools powered by these systems. Bigger models often meant broader knowledge, stronger reasoning and the ability to perform a wider range of tasks.
However, as enterprises move from experimentation to production, priorities are beginning to change. Instead of asking which model is the most powerful, businesses are increasingly asking which model is the most practical for a specific task.
According to Vijayant Rai, managing director, India, Snowflake, Small Language Models (SLMs) offer “clear technical and operational advantages,” including faster inference, lower infrastructure requirements and easier customisation for domain-specific tasks.
That shift is reflected in EY India’s The AIdea of India: Outlook 2026, which argues that while the global AI race continues to focus on trillion-parameter models, India has an opportunity to build and deploy smaller, domain-specific language models that better suit enterprise requirements. The report describes SLMs as “faster, cheaper and tailored for Indian languages and edge deployment,” highlighting their potential for enterprise AI.
Small Language Models vs Large Language Models
Small Language Models (SLMs) and Large Language Models (LLMs) belong to the same family of generative AI systems. They are trained on large datasets to understand and generate text, answer questions and assist with business tasks. The main difference lies in scale and purpose.
LLMs, including frontier models developed by major AI companies, are designed as general-purpose systems. They are trained on vast datasets, require substantial computing infrastructure and can perform a wide range of tasks without extensive customisation.
SLMs, by contrast, are designed for narrower objectives. Rather than attempting to solve every possible problem, they are trained or fine-tuned for specific industries, enterprise workflows or languages. Their smaller size generally requires less computing power, making them easier to deploy on private infrastructure or edge devices.
The distinction is becoming increasingly relevant because enterprise AI requirements differ from consumer AI. A customer asking a chatbot to write a poem has different expectations from a bank processing confidential financial information or a manufacturer analysing production data.
Why enterprises are moving beyond frontier models
The latest EY report suggests Indian enterprises are rapidly moving beyond pilot projects. According to its survey of more than 200 enterprise leaders, 47 per cent of organisations now have multiple GenAI use cases in production, compared with a much earlier stage of adoption in the previous edition of the survey. Another 10 per cent are scaling AI use cases across the business, while 76 per cent believe GenAI will have a significant impact on their organisations.
As businesses move AI into day-to-day operations, infrastructure costs, governance and deployment become more important than benchmark performance.
The report also found that 91 per cent of respondents identified deployment speed as the biggest factor influencing AI buying decisions. This reflects a shift from experimenting with the latest models to deploying AI systems that can be integrated into existing workflows quickly.
In this context, SLMs become relevant because they can often be customised for individual business functions instead of serving as universal assistants.
Why smaller AI models appeal to enterprises
The EY report argues that India’s opportunity does not necessarily lie in competing directly with global frontier models. Instead, it says the country can focus on building models that address local enterprise requirements.
As the report states: “While the world chases trillion-parameter models, India is betting on small language models that are faster, cheaper and tailored for Indian languages and edge deployment.” It adds that these models “bridge the digital divide, power vernacular chatbots and enable compliance-heavy sectors like banking and healthcare.”
For enterprises, this changes the economics of AI adoption. A retail company deploying an internal inventory assistant, for example, may not require a frontier model trained on the entire public internet. Instead, it may benefit more from a compact model trained on inventory records, supplier documentation and internal policies.
Cost, infrastructure and privacy shape enterprise AI
As enterprises move from AI pilots to production, cost is becoming a key consideration. Beyond the expense of training large models, running them for everyday tasks — known as inference — also requires significant computing resources. SLMs are emerging as an option for routine, domain-specific workloads because they use less compute while complementing frontier LLMs for more complex tasks.
Key factors driving adoption include:
-
Lower inference costs through reduced computing requirements.
-
Faster performance for specialised enterprise applications.
-
Lower infrastructure costs because of reduced hardware needs.
-
Better suitability for narrow business functions where broad general knowledge is unnecessary.
Alongside cost, data privacy and regulatory compliance are increasingly shaping enterprise AI strategies. Organisations handling sensitive information are looking for AI models that can run within their own infrastructure while meeting governance requirements.
The main factors influencing deployment include:
-
Keeping sensitive enterprise data within the organisation.
-
Meeting sector-specific regulatory requirements through on-premises deployments.
-
Strengthening AI-ready data, security and governance frameworks.
-
Hybrid cloud adoption. The EY survey found that 71 per cent of organisations prefer hybrid cloud deployments, balancing scalability with data sovereignty and governance.
Frontier models will continue to play a key role
Growing interest in SLMs should not be interpreted as the end of frontier LLMs. Large models continue to outperform smaller systems in complex reasoning, coding, research and multimodal applications. They also remain central to developing new AI capabilities.
The EY report says frontier models have made significant progress but continue to face challenges around “accuracy and enterprise readiness,” requiring appropriate human supervision.
“Most enterprises will not rely on a single model type. At Snowflake, we support a broad ecosystem of leading open and proprietary models because customers benefit from choice,” Rai said.
“The optimal approach is intelligent orchestration: routing routine or less demanding tasks to SLMs while reserving powerful frontier models for more complex needs. This hybrid model strategy offers agility, optimises costs and allows organisations to match the right AI capability to each workload as requirements evolve. Ultimately, successful AI adoption is grounded in flexibility, trusted data foundations and the ability to orchestrate multiple models for diverse business challenges,” he said.
This suggests enterprises are unlikely to choose between SLMs and LLMs in absolute terms. Instead, organisations are expected to build AI architectures in which different models serve different purposes, with frontier models handling broad reasoning and smaller models managing specialised internal workloads.
Enterprise AI is becoming task-specific
The debate over AI models is gradually moving away from the idea that bigger is always better.
As generative AI becomes part of mainstream enterprise operations, organisations are increasingly evaluating models based on deployment speed, governance, infrastructure costs and their suitability for specific business functions.
The EY report does not suggest frontier models will lose relevance. Instead, it presents a more nuanced picture in which enterprises deploy different classes of AI models depending on the task.
For Indian businesses, the next phase of AI adoption may depend less on accessing the world’s largest models and more on identifying where smaller, specialised models can deliver measurable value. If the first phase of the AI race was defined by model size, the next phase may be defined by how effectively organisations match AI systems to business problems.
