Like large language models (LLMs), SLMs can generate human-like language but are trained on smaller datasets with fewer parameters. They are said to be easier to train and use, consuming less computational power, more cost-effective, and better suited for specific tasks.
The year 2024 saw launches of a slew of lightweight models, from Microsoft’s Phi family of SLMs to Google’s Gemma and a smaller variant of Meta’s Llama model.
“Throughout 2024, large language models have pushed the boundaries of accuracy across various AI tasks, while small language models have driven mass adoption and true democratisation of artificial intelligence,” said Sundar Srinivasan, president, AI and search, at Microsoft India Development Centre.
The highly accurate, and low hallucinatory nature of SLMs makes them immediately useful for privacy-sensitive and critical sectors like healthcare and banking or finance, which are poised to see increased adoption in 2025, Srinivasan said.
Discover the stories of your interest
In healthcare, they can significantly enhance patient interaction and support, especially in regions lacking medical experts, he added. Use cases include transcribing patient interactions, data entry for electronic health records and preliminary diagnostic support.In banking or finance, SLMs will aid with personalised financial advice, fraud detection and document analysis and processing, he said.
In the coming year, SLMs will take centre stage, driven by the need for LLMs to be commercially viable for scale and their tooling becoming more developer-centric for fine-tuning them for specific needs and use cases, said Vishal Chahal, vice president at IBM India Software Labs.
Open-source initiatives in 2024 were a promising development enabling developers to fine-tune SLMs using LLMs, he noted.
“2025 and beyond will see SLMs becoming embedded into business processes and also gaining ability to run on edge devices and on-prem infrastructure, giving users control over how data exchanges with these technologies can be user controlled,” Chahal said.
Further, they will become an ideal choice for real-time GenAI applications on mobile, internet of things and edge devices, which have limited computational resources, as well as specific customer-centric tasks and personalised support, he added.
Experts said these smaller models can be expected to power more personalised digital agents and assistants for tailored experiences and responses. For instance, customer support will see exponential improvements in personalisation, efficiency, customer empathy and management of language diversity with virtual SLM-based deployments.
Legal and manufacturing sectors also stand to benefit significantly from SLM deployment, they said.
Meanwhile, the use of LLMs will be more focused on complex tasks with a need for multi-dimensional understanding across varied areas, with higher adoption in knowledge discovery and pattern mining that have a need for newer insights on large volumes of data.
