At Aveni, our mission is to transform how financial institutions apply AI responsibly, transparently, and with domain precision. That is why we are developing FinLLM, the UK’s first large language model built specifically for the financial services sector.

We are not doing this alone. FinLLM is being developed in partnership with two of the UK’s leading financial institutions – Lloyds Banking Group and Nationwide – as well as the University of Edinburgh, a global centre of excellence in AI research. This unique collaboration combines deep financial domain expertise with cutting-edge language modelling and responsible AI design.

Where general-purpose models fall short, lacking context, control, and transparency, FinLLM is set to solve this. It is built from the ground up for the workflows, regulations, and language of financial services. Aveni Labs, Aveni’s Research and Development department, has spent the last 6 months developing and rigorously testing FinLLM – the results speak for themselves.

What follows are the key technical findings that demonstrate how FinLLM delivers superior real-world performance across financial services tasks, unlocking innovation while managing the operational and regulatory demands of enterprise AI.

1. Designed for your regulatory reality

Generic LLMs are not built for FS. FinLLM was. It was developed specifically to align with the FCA, PRA and EU AI Act guidelines, addressing the sector’s highest standards around data privacy, auditability, and responsible AI deployment. This means you can build with confidence knowing the model’s design principles are rooted in regulatory compliance from day one.

2. Superior performance on financial tasks

FinLLM consistently outperforms leading general-purpose and open-source LLMs (e.g., Gemini 1.5 Flash and GPT-4o mini, LLaMA, and Mistral) across the kinds of tasks financial firms care most about:

• Text classification, such as flagging conduct risk in adviser calls
• Long-context reasoning, used in reviewing policy documents or investment reports
• Tabular data analysis, like interpreting balance sheets or onboarding forms
• Multi-turn dialogue modelling is essential for AI agents that hold coherent, compliant customer conversations

These capabilities were tested using AveniBench – FinLLM’s in-house evaluation suite.

→ AveniBench is a benchmark dataset built from real financial services tasks, such as compliance monitoring, customer vulnerability detection, and KYC document review. It ensures FinLLM is not just accurate in lab conditions, but relevant in production.

3. Tailored, transparent and tunable

FinLLM gives your teams direct control over the model:

• Open model variants (1B and 7B parameters) that can be securely deployed and fine-tuned on-prem or in Virtual Private Cloud (VPC) environments
• Training tuned for financial communication, so the model understands nuance in situations such as product disclosures, customer complaints, and suitability discussions
• Model transparency built in, so outputs can be interrogated, audited, and explained to regulators or internal risk teams

This is made possible by innovations like Finance Classifier 2.0, a proprietary component developed by Aveni that classifies whether content is truly ‘financial’.

→ Finance Classifier 2.0 is a filtering system used during training to ensure the data the model learns from is genuinely financial. It weeds out noisy or irrelevant text (e.g. marketing jargon, outdated forum chatter) and prioritises high-quality sources like regulatory guidance, financial advice transcripts, and product literature.

4. Safe, ethical and future-proof

FinLLM’s architecture supports safer, more accountable GenAI, essential in high-stakes domains like finance as well as in use within agentic frameworks where autonomous action is involved versus just information generation.

Its training process incorporates safeguards, for example, all data included in the training of FinLLM undergoes filtering to flag sensitive, toxic and biased data before the model even sees a single token, and future releases will include input/output guardrails and response justifications to meet growing regulatory expectations.

One of the most exciting developments is FinLLM’s integration into agentic RAG systems.

→ Agentic Retrieval-Augmented Generation (RAG) means combining FinLLM’s language capabilities with retrieval engines that fetch facts and context from real-time document sources, like policies, case logs, or regulatory guidance. This allows the AI to cite sources, answer accurately, and operate with traceability, giving teams confidence in outputs that are both informed and auditable.

5. Built with the industry, for the industry

FinLLM was not built in a vacuum. It was shaped in collaboration with Lloyds Banking Group, Nationwide, and the University of Edinburgh, ensuring every design decision was informed by live, high-priority problems in financial services, with use cases from scaling call QA to reducing the burden of regulatory reporting.

Turning FinLLM into business value

FinLLM is more than a model, it is a critical driver of AI infrastructure, designed for the complexity, compliance, and scale of UK financial services. Whether you are modernising operations, building next-gen products, or accelerating regulatory transformation, FinLLM provides the foundation you need.

Here is how forward-thinking institutions could leverage it:

• AI-powered customer interaction analysis: automate the review of adviser-client conversations to identify vulnerability, conduct risk, and service gaps, at scale and with confidence.
• Next-gen digital assistants: deploy domain-specialist AI agents to support customers, advisers, and ops teams with compliant, context-aware guidance across financial products and policies.
• Intelligent document understanding: extract insights from lengthy, complex documents like suitability letters, mortgage offers, or KYC forms, without human review bottlenecks.
• Faster, safer complaint resolution: classify, triage and summarise complaints with high accuracy ultimately reducing response times while improving consistency, defensibility and oversight.
• Risk and compliance monitoring at scale: continuously scan communications, documents and decisions for regulatory red flags, proactively identifying issues before they escalate.
• Accelerated product development: equip product and engineering teams with a fine-tuned LLM that understands financial workflows, speeding up prototyping and safe deployment of AI-powered services.
• Policy-aware automation: implement AI systems that operate within the boundaries of your internal policies and risk appetite, with explainability and auditability built in.

FinLLM marks a turning point for financial AI. It is not just a model, it is a catalyst for change. Purpose-built for the language, regulation, and complexity of the industry, it empowers financial institutions to reimagine what is possible with GenAI. For leaders ready to shape the future, FinLLM is the critical foundation.

Comparison of Aveni’s  fine-tuned FinLLM 7b model against frontier general-purpose models

Read the original article here.