The Financial Conduct Authority has said the results of its small language model for financial services with Malted AI, the Edinburgh-based startup, had “extremely encouraging” results.

Edmund Towers, head of advanced analytics and data science units at the FCA, detailed the experiment, which began in 2025 with Malted AI, the bespoke AI language model firm, in a blog on Linkedin.

The experiment was led by Towers and Alister Shepherd, chief scientific adviser and CISO at the FCA, and aimed to work with Malted AI to see where small language models (SLMs) can add value to the FCA and financial services, as well as any tradeoffs SLMs have versus large language models.

The focus on SLMs is largely due to LLM’s comparatively large costs and compute requirements, where as SLMs use a fraction of these required parameters.

SLMs also offer specialist performances tailored to specific datasets, with lower energy and infrastructure requirements and costs, fast inference, and greater auditability.

The FCA partnered with Malted AI to create a SLM prototype for the high-volume task of multilabel classification, which deals with a range of issues from vulnerability detection to customer routing.

The prototype used a 400 million parameter model and was trained on 10,000 records, running on a single tenant environment with no data sub-processors. This provided the researchers with full control over the data. Further, the experiment allowed for precision-recall threshold calibration to match potential risk.

The SLM results showed “zero missed high risk classifications, alongside materially higher accuracy than a prompted general LLM on the same task,” Malted CEO Iain Mackie said. “We also observed significantly faster latency, with large reductions in compute cost and energy usage for the use cases in scope.”

Towers wrote: “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.”

Recommended reading

• Report: Top Banks Advancing AI 2.3x Faster Than the Rest
• FCA Launches Review On AI In Financial Markets
• Finance Leaders See 2026 As the Year to Scale AI

SLMs were viewed as being capable of delivering speed and strong data control benefits for the right use cases, while LLMs were still the best choice for more complex generative operations and open-ended reasoning.

“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,” Ewen Fleming, head of financial services at Malted said.

While the experiment showed the promise of SLM, Towers emphasised that they should be seen “as an important component within a broader AI strategy, rather than a replacement for general purpose models.”