Small Language Models (SLMs) and Your Data Services Provider

You may have heard about Small Language Models (SLM) recently and wondered what those are and how they fit into the AI developments for business. You may also have considered whether that’s something for your organization as you progress along the GenAI or Agentic AI path. There are distinct differences between SLMs and LLMs, not just in size, but also in cost, application, and resource usage.

At DATAmundi, we’re constantly doing research on how to best use data to fuel cutting-edge AI technologies. Below, we lay out the benefits of SLMs and how they might be better suited to agentic AI applications as well.

What is a Small Language Model?

A small language model is a subset of language models that perform domain-specific applications using fewer resources than large language models (LLMs). They are designed for specific tasks, trained with fewer parameters, and are optimized to run on devices with limited computer power. They often excel at niche applications, such as on-device mobile assistants.

Some of the top SLMs are Apple’s OpenELM, Microsoft’s Phi3 mini, or Google’s Gemma, among others.

An example of how a business might use an SLM over an LLM can be found in a recent DATAmundi case study, where we sourced data for a client building an on-device assistant specifically for people with vision impairments. A very niche, on-device application.

How does it work?

Small language model benchmarks show that SLMs – generally operating below 10B parameters – achieve performance comparable to LLMs in commonsense reasoning, instruction following, tool calling, and code generation.

SLMs are proving to be an operationally better fit for agentic systems. Agentic tasks are often repetitive, narrow, and structured, making large generalist LLMs excessive and wasteful to use. With agentic AI naturally allowing for heterogeneous architectures, it’s possible to use SLMs for specialized tasks and use an LLM only for complex reasoning.

With a smaller language model, you’re also looking at a smaller overall footprint, so lower operating costs and fewer compute resources. They can run on-device, which improves privacy control, and can be integrated into multiple pipelines for different tasks.

To make SLMs successful, it’s important to train them on very well-sourced and annotated data. Specific applications require specific – niche – data. Part of that is also making sure you remove any biases that can easily slip in when sourcing niche data.

How does DATAmundi fit in?

To make sure your SLM is used to its fullest extent and hitting the benchmarks you want, you’ll want to train it on expertly annotated data that’s been ethically sourced and free of biases. That is, of course, easier said than done.

This is where DATAmundi excels. We can source, annotate, and fine-tune your SLMs for multiple languages and across cultures. In past projects, we’ve sourced experts who were both linguists and subject matter experts in a specific – niche – field so we could deliver clean and clear data in multiple languages. Our clients used that data to train their AI for customer service chatbots, digital assistance, internal risk evaluation for business use cases, and automation of reference validation for scientific articles.

As the need for your SLM application grows, we have the expertise to help you scale up. Generally, SLMs are a good way to step into Agentic AI without having to make large investments in an LLM infrastructure. LLMs remain important for large and complex reasoning, but we tend to see that the future of scalable AI agents is in specialized small models.

Take a look at some of our case studies, which will show you that we know how to prepare your data and optimize your AI for both niche projects and larger projects.

At DATAmundi, we offer: