When Size Really Matters: Small Is the New Big in AI
For the past few years, the artificial intelligence industry has been obsessed with size. The bigger the model, the more it can do — and even more interestingly, the more emergent properties it might reveal. Many of the capabilities we now take for granted — from writing to creativity — were emergent behaviors that surprised both developers and researchers. Today, some of the latest models have passed the two-trillion-parameter mark, making them incredibly powerful… and incredibly expensive to build and run.
But now there’s a new kind of model that isn’t competing to be the biggest, but rather to be more certain about what it knows (and what it doesn’t). Allow me to introduce you to Small Language Models, or SLMs. Unlike the massive systems we rely on for almost everything, SLMs are smaller, more tightly tuned, and trained for specific use cases: health care, finance, internal company knowledge, supply chains — you name it. The catch? There’s always a catch. They won’t perform as well on general tasks, which means you’ll still want to keep your favorite LLM handy.
At this point, you might be thinking: “Why would I want a smaller AI that’s only good at one thing?” After all, you’re someone who wants the full package, right? Well, there are several compelling reasons to take this new generation of models seriously.
The Upside of Being Small
First, when developed correctly, SLMs can outperform traditional LLMs in highly specialized tasks. How? By training them on tightly curated datasets and applying fine-tuning for very specific functions. Because they’re focused, free of irrelevant information and noise, they have fewer opportunities to deliver an incorrect answer. And when it comes to your work, you’d probably prefer a system packed with knowledge relevant to your field rather than an LLM trained on a huge blend of public and private data that, let’s be honest, likely has very little to do with your day-to-day reality.
Second, as with any emerging technology, once we’ve proven that something can work at scale, the next natural step is to optimize it — reduce operating costs and minimize environmental impact. One of the major advantages of SLMs is that, by being smaller, they consume far fewer resources: less compute, less electricity, less water. For context, a single interaction with a popular generative AI model can use up to ten times more energy than a Google search and consume around 50 ml of water. SLMs, in contrast, are far more efficient and operate within much more reasonable limits.
Third, their compact nature allows them to run directly on your phone, laptop, or tablet without needing to send information to the cloud. Keeping AI on your own device gives you greater control and security, allowing organizations and individuals to process sensitive information without exposing it. This local-first capability — no cloud needed — not only expands where the models can be used, but also strengthens security. Picture working with your customized AI safely on a flight, without having to pay for (or gamble on) the in-flight wifi.
Ready to Put SLMs to the Test?
It’s worth noting that SLMs aren’t fully mainstream yet. But I know my readers like to stay ahead of the curve, and all signs point to consumer- and enterprise-grade offerings arriving sooner than most people expect. If you have an internal AI team, you can get started today… after forwarding this article to your CTO and colleagues. And if you’re wondering how these models might fit into your organization, here are four key questions you and your team should consider:
What kinds of privacy trade-offs have we considered when using AI in our organization?
Is there a clear need for our teams to access AI directly from their own devices?
Would a specialized AI trained for specific use cases give us a meaningful advantage over general-purpose models?
Do we have an internal or public sustainability mandate?
If you answered yes to any of these, then you already have a solid reason to consider making small the new big at your company. So… is small the new big? It depends. The real shift isn’t choosing between big and small — it’s understanding which one serves each situation best.
In the years ahead, the companies that truly thrive will be those that stop treating AI as a monolith and start seeing it as a toolkit. That means deploying the right model, at the right scale, for the right task. There will be plenty of situations where a specialized SLM is the best choice for your organization or even your personal workflow. The answer will ultimately depend on the AI strategy you’ve built and how SLMs, LLMs, agentic AI, and other types of systems align with your goals.
Don’t worry — my next piece will cover best practices for designing your AI strategy, because the future of AI won’t be defined by size, but by how well it aligns with what you need. And when that alignment is right, small really can be the new big.
Originally published in Spanish for Fast Company Mexico:
https://fastcompany.mx/2025/09/22/slm-small-language-models-ia-tamano/
