Small language models (SLMs) could hit the mainstream in 2025, according to analysts, as enterprises look to speed up training times, lower carbon emissions, and bolster security.
While much of the generative AI boom has focused on LLMs and an industry arms race to create more powerful models, Isabel Al-Dhahir, principal analyst at GlobalData, believes the appeal of leaner options will surge in the year ahead.
A key factor in this prediction is the fact that SLMs use smaller and more focused datasets, enabling enterprises to train models in a matter of weeks rather than months.
SLMs typically boast fewer than 10 billion parameters, for example, compared to up to a trillion for larger models.
“The use of focused datasets makes SLMs particularly well-suited to domain-specific functions and small-scale applications such as mobile applications, edge computing, and environments with limited compute resources,” she said.
“As training techniques improve, SLMs with fewer parameters are becoming more accurate and can have a much faster processing time.”
These smaller datasets also make SLMs more resilient from a cybersecurity perspective, Al-Dhahir noted, as they represent a smaller attack surface and can be operated locally with relative ease compared to larger, more cumbersome models.
By definition, SLMs are less expensive and energy-intensive to run, as they need far less computing power than LLMs; nor do they need expensive infrastructure.
Another advantage is that it’s easier for them to meet regulatory requirements. Not only is it more straightforward to obtain licenses for training material, but they avoid stringent obligations as they don’t meet the computing threshold.
SLMs are gaining traction
A host of major industry players are working to offer SLMs, with Microsoft, Meta, and Google having all recently released their own models.
Microsoft, for example, is offering the Phi-3 family of small language models, aimed at the creation of marketing or sales content, along with customer support chatbots.
Earlier this year, Google launched Gemma 2B and Gemma 7B. These two model iterations assist with text-generation tasks like answering questions and summarizing information.
Meanwhile, Mistral, has published one of its models under the Apache 2.0 license.
SLMs do have their limitations, however. There can be a tendency for organizations to find SLMs effective enough to start with, but then later need to transition to larger models.
And because SLMs are often built to excel in specific areas, it can be tricky to get them to work with the models outside their remit.
Earlier this year, Arun Subramaniyan, founder and CEO of Articul8, told ITPro that SLMs are best used in tandem with LLMs – a view with which Al-Dhahir agreed.
“SLMs are not meant to replace LLMs but rather complement them. There is still a lot of appetite for the capabilities of generative AI, and many organisations are still getting to grips with where it can serve them best,” she said.
“As competition in the AI market intensifies, companies are under increasing pressure to show a strong business case with demonstrable ROI. SLMs, with their suitability for industry-specific applications, offer easier scalability across diverse environments.”
