Press
Exploring Small Language Models at the FCA
Edmund Towers, Head of Advanced Analytics & Data Science Units, FCA
Supporting innovation is central to the FCA’s strategy, and AI is a key part of that journey. While we help firms adopt AI safely through initiatives like AI Live testing and our enhanced sandbox, we’re also exploring how emerging technologies can strengthen our own work. This helps us to be a smarter regulator – embracing technology so we can be more efficient and effective. In turn, this helps us deliver for consumers, markets and the wider economy.
Recently, myself and Alister Shepherd had the chance to lead a practical experiment with Malted, a UK based startup building small language models (SLMs) for financial services. Our aim was simple: understand where SLMs can add real value and what tradeoffs they involve compared with large language models (LLMs).
Why look at small language models?
LLMs have become the dominate focus of AI adoption across financial services - and for good reason. They’re powerful generalists that perform well across a broad range of tasks. But they’re also large, compute intensive and costly.
SLMs, by contrast, are a fast growing area of interest across UK financial services. As their name suggests, they are much smaller - typically operating in the hundreds of millions to a few billion parameters - and can offer:
Strong specialist performance when fine tuned on domain specific data
Lower energy use and infrastructure requirements
Greater transparency and auditability, thanks to more common open weight availability
Lower costs for targeted use cases
Very fast inference, supporting real-time workflows
Both model types have a place. Our goal was to understand where SLMs could excel in practice in financial services.
Testing SLMs with Malted
In summer 2025, we partnered with Malted to prototype an SLM for multilabel classification - a high volume task across financial services, from customer routing to KYC/KYB, complaint triage, prioritisation and vulnerability detection.
Using a 400 million parameter model trained on around 10,000 records, Malted built a classifier tailored to our specific context. The model ran in a single tenant environment with no data sub-processors, giving us full control over data handling. We also calibrated precision–recall thresholds to reflect operational risk appetite.
The results were extremely encouraging. As Malted’s CEO, Iain Mackie, PhD , highlighted, the SLM achieved “zero missed high risk classifications, alongside materially higher accuracy than a prompted general LLM on the same task. We also observed significantly faster latency, with large reductions in compute cost and energy usage for the use cases in scope.”
This proof of concept reflects the FCA’s commitment to innovation that is grounded in responsible choices. As AI continues to evolve rapidly, it is essential to understand how technical design decisions connect directly to risk appetite, transparency, and trust
We're increasingly seeing UK firms adopt a blended approach focused on use case return on investment - using large models for general workflows, while deploying smaller, specialised models where performance, privacy, and scale really matter.
What this means for financial services
Our findings reinforce a simple message: SLMs can deliver excellent performance for the right use cases, alongside efficiency, speed and strong data control benefits. LLMs, meanwhile, remain the best option for complex generative tasks, RAG workflows and open-ended reasoning.
As Ewen Fleming , Head of Financial Services at Malted, reflected: “We’re increasingly seeing UK firms adopt a blended approach focused on use case return on investment - using large models for general workflows, while deploying smaller, specialised models where performance, privacy, and scale really matter.”
Our work with Malted highlighted that SLMs can deliver exceptional speed and efficiency for specialised tasks. They should be viewed as an important component within a broader AI strategy, rather than a replacement for general purpose models
Looking ahead
This project is one part of the FCA’s wider programme of technology and AI adoption. Alongside our work to deploy AI systems into production we are also aware that the technology landscape is continually evolving , we’ll continue testing emerging tools, partnering with UK fintechs and sharing what we learn about what responsible and effective AI deployment looks like in practice.