New Delhi: It was roughly four years back that ChatGPT, Claude and Gemini rode the AI revolution wave with their large language models or LLMs, doing everything from summarising large reports to preparing diet plans and therapising a user in need.
Large language models work on a simple principle: they predict the response to your query based on the enormous amount of text they have read. The more data used to train a model, the more relevant its response. This is what makes the chatbot interface so easy; it’s almost like talking to a very bright friend. However, LLMs require massive amounts of electricity and water for computing.
“Running a 400-billion-parameter model to perform a simple intent classification is like using a Boeing 747 to deliver a pizza,” wrote AI expert and angel investor Rob May in his Substack article.
As AI evolves from answering in chatbots to acting as agents, there is a need to look at more viable models that are environmentally friendly.
This is where Small Language Models or SLMs come into the picture. Unlike LLMs, SLMs are tailor-made for custom tasks and generate responses based on a more focused, narrower dataset. As a result, they require less energy to operate and are more cost-effective.
Let’s illustrate this with an example. Say you need legal advice on labour codes. An LLM will act like a senior consultant lawyer who has experience in everything from corporate law to intellectual property to international law. Whereas an SLM will act like a lawyer specialising in employment law. SLMs are not meant to handle everything, but they may outperform a generalist on that specific subject.
As the industry moves toward agentic workflows, where the AI model is not just chatting with you but also making decisions and performing a series of tasks on your behalf, requirements change. One doesn’t need a model that can deduce Shakespearean sonnets and also write code. But one needs a model that can stick to writing code and debug errors while using less energy.
In a paper titled Small Language Models are the Future of Agentic AI, published by NVIDIA Research in 2025, researchers argued that using one large general-purpose model for agentic workflow is wasteful and that heterogeneous systems using SLMs are more efficient.
“We argue that insisting on LLMs for all such tasks reflects a misallocation of computational resources—one that is economically inefficient and environmentally unsustainable at scale,” the study noted.
The NVIDIA study also pointed to a “10-fold discrepancy” between the market for LLM-driven AI services estimated at $5.6 billion in 2024, and investment in cloud infrastructure, which reached $57 billion. In such a situation, efficiency is the only path to profitability.
When compared to LLMs, SLMs are faster, require less Video Random Access Memory (VRAM) and consequently cheaper hardware, and are more accurate. They can even outperform models 10 times their size. That said, SLMs are designed to complement, not replace, LLMs.
While LLMs are better for broader unfamiliar tasks, SLMs work better in use-cases where the workflow and tasks are clearly defined — such as translation, citizen service queries, or clinical triages.
“LLMs are very powerful and we cannot ignore them completely. There are many tasks that LLMs can do but SLMs can’t. For example, detecting vulnerabilities in a cybersecurity system, which is a more strategic task, can’t be performed by SLMs,” said Manindra Agrawal, Director, IIT Kanpur.
Cheaper and greener
SLMs, like Microsoft Phi-2 and Google Gemini Nano, are significantly smaller in size compared to their large counterparts. Fine-tuned SLMs are particularly well-suited for industry-specific applications in legal, medical and financial services.
According to a 2024-article published by UNESCO, SLMs can wield substantial influence in higher education, serving as subject-specific generative AI tools. It also indicated that SLMs offer a more economical alternative to countries with limited digital infrastructure and resources.
“Although the Global North continues to dominate upstream technologies and platforms, Global South countries can use SLMs to cultivate home-grown AI tools and foster supportive local ecosystems,” the article observed.
Training frontier‑scale LLMs — which are the most advanced and built on hundreds of billions of parameters — requires enormous electricity as well as water for cooling of data centres, in addition to vast land for infrastructure. For climate-vulnerable countries, particularly those that are not equipped to absorb the shock, SLMs provide a greener alternative for smaller problems.
Not every country needs to build its own large — trillion-parameter— model to survive the AI era and avoid being crushed by technological Darwinism.
“A significant proportion of countries have datasets that are subcritical for training frontier-scale general-purpose models, but those datasets may still be entirely sufficient to build, adapt or ground smaller, domain-specific systems,” said Prachetas Bhatnagar, a GovAI expert specialising in applied AI, tech policy and government transformation programmes, who leads Strategy at the Tony Blair Institute for Global Change in London.
Bhatnagar proposed what he named as an “F-minus strategy” to provide a practical framework for this. The strategy requires governments to determine the minimum level of model capability required for a defined task and then optimise for cost, auditability, linguistic relevance, reliability and ease of deployment.
“In many cases, an F-minus-5 or F-minus-6 model, built around nationally relevant data and institutional workflows, could deliver substantial efficiency gains across public services, improve the quality and reach of government delivery, and create entirely new forms of public value without requiring frontier-scale infrastructure,” Bhatnagar said.
For instance, if the Indian government wants to digitise a paper-based workflow, say in the Income Tax department, an SLM trained on India’s taxation rules is better suited for the use case, as compared to a frontier, general-purpose model. Depending on the nature of queries, a constellation of interconnected SLMs can also be an effective model for this use case.
Also read: AI model detects hidden health risks using standard sleep study data
Geopolitically smarter pick
Home-grown SLM tools can also solve another problem, which inherently creeps into LLMs operating on Western datasets. Most existing LLMs are trained on what psychologists and anthropologists call WEIRD or Western, Educated, Industrialised, Rich, and Democratic.
Dominant LLM models trained on WEIRD data reflect the values of developed economies and end up misinterpreting the cultural nuance and indigenous knowledge of emerging economies. The outcome is more serious than expected. When generative models label countries as high‑risk or low‑growth based on biased data, it influences investor sentiment as well as government priorities.
“We do not know what has gone into training existing LLMs, which are predominantly developed by the US or China. Giving a hidden sequence of a 100-letter command may trigger the model into behaving erratically. We can’t rely on these models when it comes to use cases like security. Hence, both LLMs and SLMs need to be developed indigenously,” Agrawal said.
Since SLMs are cheaper to build, train and fine-tune, organisations can create models tailored to local needs—such as indigenous languages, dialects, legal systems, and public services—instead of relying on one-size-fits-all global models. This localisation makes the outputs more relevant and culturally appropriate. SLMs can grant a level of data sovereignty to countries, particularly those that face resource constraints. In that, SLMs allow them to define their knowledge, train the models in that knowledge, and control their narrative and technological priorities.
For instance, if one is developing an AI model that provides real-time agricultural advisories to farmers in rural India, the model must be trained in regional Indian dialects and local knowledge for it to be effective.
The Principal Scientific Adviser (PSA) to the Indian government released a white paper titled Advancing Indigenous Foundation Models in March this year. It also urged India to build sovereign AI models that can cater to the country’s linguistic diversity, strategic priorities, and public-sector needs, while reducing dependence on foreign foundation models.
Last year, the Indian government selected Bengaluru-based startup Sarvam AI to build India’s sovereign large language model under the IndiaAI mission. But the startup has also rolled out smaller, highly optimised models—Sarvam 1, Sarvam-M and Sarvam Edge—which are trained on indigenous languages and perform specific tasks including translation, text-to-speech conversion and speech recognition.
“India has a few SLM models that are being trained and deployed on a pilot basis. They have not yet been deployed at scale. More research needs to be done on developing indigenous SLMs that do not compromise India’s digital sovereignty,” Agrawal said.
(Edited by Ratan Priya)
This is insightful!