Skip to Content

Unlocking the potential of small language models (SLM)

Nisheeth Srivastava
Nov 6, 2023

Its fashionable to hear a lot of folk berate LLMs (Large Language Models) these days for what they are or represent.

I’m a strong believer in #genai tech at large, while recognizing that it is in its infancy and it will morph into something altogether different as time goes by. This current imperfect stage of LLMs is crucial though, as the world plays with it just so that future relevance is woven into what was an originally a mere experiment to predict the next word in Amazon reviews online (or so the story goes I think!).

Yes LLMs or GenAI or whatever are not perfect, no you can’t use them for everything conceivable, but hell yeah you can for countless use-cases still! Frankly, I observe a heightening of awareness in my own teams on contextualizing so many disparate ideas, thoughts and themes to business priorities, which are complex to grapple with as one works with top business leadership and teams globally, trying to come up with something astounding and impactful in their journey of future relevance. And we’ve made remarkable strides.

That said, I do believe that Small Language Models (#SLMs) are ones to keep one’s eye out for. As I build one of these in my Innovation labs to play around with, I get quite taken by the possibilities, and in my search for analogies to support my point of view, I was drawn to the phenomenon in recent times of Vertical SaaS companies doing rather well. Actually, FAR better than horizontal SaaS ones. Take a look at the visual and I hope you can draw your own inferences. As for mine, they’re as follows:

1. SLMs are a cake-walk when it comes to training them. More importantly, they’re EXPLAINABLE (if you want them to be!).
2. SLMs are more customizable with their fewer parameters and targeted trainings datasets.
3. Leaner with lower carbon emissions, as the count of neuralnet synpases to cover in compute cycles are far fewer.


I mean… if all I want to do is create creative marketing content for my brand or print product descriptions online, or engage with standard customer queries in non-regulated industries, or provide inferences and NBA on your commercial/ supplier contracts, or drive some automation, what good is it to use a trained model on all things www?!

Vertical SaaS companies are doing much better than the horizontal ones, follow a similar principle, and I do believe all these #startups and #scaleups will do well to bring in the sheer power of SLMs into their product stack. As will I :-)

Figures speak for themselves. With due thanks to Allied Advisors for this report.

Unlocking the Potential of Small Language Models

Meet the author

Nisheeth Srivastava

Chief Technology and Innovation Officer – India, Capgemini