Every research project begins with a decision about what matters. For Aveni Labs, FinLLM began with a simple line in the sand: if AI is going to sit inside UK financial services, it needs to understand that world on its own terms. Over the last twelve months, that idea turned into a full model family, trained on more than 91 billion carefully selected tokens, evaluated on an in-house benchmark built from real financial tasks, and deployed into production workflows that advisers and reviewers use every day.
This overview walks through that journey from a lab perspective. It covers how we built AveniVault as a financial pretraining corpus, how AveniBlocks grew from fifty thousand to a quarter of a million supervised examples, and how we moved from a 1B prototype to 7B, 14B and 24B models that outperform strong public baselines on finance-specific benchmarks. You will also see how synthetic data lifted tabular reasoning scores, how AVENIBENCH keeps our evaluation tied to real use cases, and how FinLLM now powers vulnerability detection and fact extraction inside Aveni Detect and Aveni Assist.
If you want to see what it takes to turn an open-source base model into a domain-specific LLM that can satisfy governance, privacy and performance requirements in UK financial services, this is the place to start. The full report goes into the details: data pipelines, training schedules, safety benchmarks and the practical lessons learned along the way.
