FinLLM is the first LLM family built for UK financial services
A suite of small language models trained, evaluated and safety-tested on UK regulatory, advisory and customer-conversation data. It powers every Aveni product.
AI that speaks the language of financial services
Generic AI models are trained on the open internet. FinLLM is trained on the data financial services works with: regulation, advice files and real customer interactions. Every model is designed for the work firms need to do, evaluated against real financial services tasks, and deployed in regulated environments.
One family. Every financial services job.
A compact model sized for real-time, widget-scale inference. Runs on a single GPU or on-device, with the speed to sit inline inside customer-facing journeys without adding latency.
Evaluated where it matters
FinLLM is evaluated on AveniBench Finance, a suite built for UK financial services tasks, and on data from live deployments.
AveniBench Finance
Average score across UK financial services tasks.
| Model | Score |
|---|---|
| FinLLM-24B-M | 57.54 |
| Gemini 2.5 Flash | 53.90 |
| GPT-4.1 Mini | 53.26 |
| FinLLM-14B-M | 51.55 |
| Mistral Small 3.1 24B | 43.01 |
Higher is better. Source: AveniBench Finance, Year of FinLLM report, Nov 2025.
Vulnerability detection
Identifying vulnerable customers in recorded adviser conversations. Missing a vulnerable customer is a regulatory failure, so catching every real case matters most.
| Model | Macro F1 |
|---|---|
| FinLLM 8B-Compliance | 0.957 |
| GPT-4o | 0.67 |
| Aveni Vulnerability Detector (prior) | 0.67 |
Higher is better.
Safety
AveniBench Safety score: resistance to toxic, biased and misaligned outputs on finance-specific prompts.
| Model | Safety |
|---|---|
| FinLLM 7B Q-Tab | 69.37 |
| FinLLM 7B Q (safety-prompted) | 67.69 |
| Qwen 2.5 7B | 63.73 |
| FinLLM 7B Q (baseline) | 62.33 |
Higher is better.
On the tasks UK financial services runs, FinLLM outperforms frontier models at a fraction of the parameters and cost.
Designed for regulated financial services
Trained on UK financial services data
Built on AveniVault, Aveni’s corpus of 91B+ tokens of UK regulatory, advisory and customer-conversation data.Evaluated on financial services tasks
AveniBench measures what matters: suitability, conduct risk, vulnerability, complaint detection, and safety.Small, specialised, deployable
Models from 3B to 24B parameters. Run on your infrastructure or ours.Safety-tested for regulated use
Red-teamed (adversarial testing) across toxicity, bias, IP infringement, privacy, misinformation, misalignment, hallucination and sustainability, aligned to FCA, PRA and EU AI Act expectations.
Aveni’s FinLLM will be a game changer for UK financial services. Recognising its potential, Lloyds Banking Group invested in Aveni in 2024, and since then have worked closely with Aveni to co-create FinLLM and test it on our live AI use cases. Having seen the FinLLM roadmap and integration with the broader Aveni product suite, I’ve been blown away with the progress and ambition. I am excited to see the transformative impact the Aveni FinLLM will have when deployed at scale across Lloyds Banking Group and the industry.
Since investing in Aveni and working closely together on co-creating the FinLLM, we are delighted to see its first iteration being released. We’re excited by the performance of FinLLM and the potential benefits it will bring both Nationwide and our customers.
Read more from the team
Trusted by leading financial institutions, delivering results in real-world environments.
5 min read
11 min read
10 min read