Aveni CEO: Why small AI models matter to UK finance

Smaller, specialist AI models are becoming a practical priority for UK financial services firms because they cost less to run, suit compliance work and reduce dependence on a few US technology providers, argued Joseph Twigg, CEO of financial services AI firm Aveni.

He added that any model bigger than the use case needs is an inefficient use of tokens, compute and energy. Deciding whether a customer has complained, for example, needs a model of a few hundred million parameters that can be hosted locally, he said, rather than one with trillions.

While major U.S. labs don't publish parameter counts, Alibaba's Qwen 3.8 Max model in China officially claims 2.4 trillion parameters, and DeepSeek's V4-Pro features 1.6 trillion total.

He said in an interview that its FinLLM models range from one billion to 24 billion parameters, the internal values a model uses to process language, compared with trillions for the largest general-purpose systems. Lloyds Banking Group and Nationwide Building Society invested in Aveni and co-developed the FinLLM model.

Aveni started in 2018, when it was building machine learning models rather than competing with today's large language models. He said the message is the same as it was then: understanding a sector's processes and accessing its data improves model performance and accuracy.

" You don't need a two-trillion parameter model to decide whether someone's complained or not. You need probably a few 100 million parameter which can be locally hosted. It's far, far more efficient. ...Within the financial services, FinLLM, we found two or three years into the post GPT era of enterprises trying to adopt AI now Token consumption is becoming a real thing," said Twigg. " The objective is not superintelligence. The objective of the large hyperscaler is superintelligence. ... The objective for any enterprise is requisite intelligence."

He expects more suppliers of smaller specialist models as enterprise AI settles. Until now, he said, most enterprises have followed the advice of Google, Microsoft, OpenAI and Anthropic.

The cost of getting this wrong is becoming visible, he said. For compliance use cases at tier-one firms handling tens of millions of interactions a year, token consumption- the volume of text a model processes quickly reaches millions of pounds, he estimated.

Aveni's models are designed for seven or eight specific compliance monitoring use cases, such as assessing whether someone said something in the right way. He said this work has proved hard for enterprises to scale with general-purpose models because it requires repeated checks and because language models can be convincing when they hallucinate.

Twigg said the emphasis in the UK on sovereign AI has grown in recent months, alongside concerns over AI safety. Firms are wary of embedding one of two or three companies at the heart of their operations, industry or country, he said, because those companies could switch services off or use them as leverage. He said the issue had become more important to financial service providers in the Trump era.

"There is a massive fear at the moment that hyperscalers would fundamentally disintermediate the market. If your financial services customers start to interact with their personal financial lives via ChatGPT or Anthropic, they own the customer. What happens to the financial services industry is they become effectively manufacturers," added Twigg.

Protecting the market would take intervention at the country, regulator and industry level, Twigg said, for long enough to allow domestic alternatives that keep revenue within the country's borders. 

In July, the UK designated AWS, Microsoft, Google Cloud and Oracle as Critical Third Parties, overseen jointly by the Bank of England, Prudential Regulation Authority (PRA) and the Financial Conduct Authority (FCA). The regime is intended to reduce the risk that a failure at one of these providers disrupts the wider UK financial system, giving regulators powers to gather information and set rules for the services they supply to financial firms.

Over the past two years, Twigg said, large enterprises have moved from buying AI to building it themselves on infrastructure from hyperscaler partners, but many internal projects have not progressed beyond pilots or proofs of concept.

"The availability of tech is not really the big challenge right now. The big challenge is that large tier one enterprises, in particular, do not have the appropriate level of expertise internally to do anything beyond what hyperscalers tell them. The operating model is primarily the challenge."