Why Finance Needs Its Own Large Language Model: Insights from Du Xiaoman

The era of large models has arrived, prompting organizations to reshape their knowledge systems and continuously learn while leveraging AI tools for efficiency. This piece shares Du Xiaoman's exploration and practice of large‑model technology in the financial sector.

1. From General Large Models to Financial Large Models

Large models exhibit unexpected capabilities that can create incremental value for finance. The industry is characterized by three traits that align closely with LLM strengths:

Data‑driven: Financial decisions rely heavily on data analysis.

Highly knowledge‑intensive: Finance contains abundant specialized terminology and complex sub‑domains such as wealth management, insurance, and credit.

Complex business processes: Numerous collaborative steps make workflows intricate.

These characteristics match the logical reasoning, memory, and understanding abilities of large models, making finance a natural application scenario.

2. Challenges of Applying General Models to Finance

General models struggle with financial tasks for three main reasons:

Domain knowledge gap: Specialized terms (e.g., KS, MOB, COB) are often unknown to generic models.

Capability issues: Hallucinations, inaccurate calculations, and forgetting undermine the high‑precision demands of finance.

Cost of application: Training and inference of massive models are prohibitively expensive.

Training a 70‑billion‑parameter LLaMA‑2 model on 2 trillion tokens requires roughly 187 days on 48 A100 GPUs, and inference for such a model needs two A100 cards (≈129 GB memory). These costs are out of reach for many enterprises.

Domain‑specific, smaller models can achieve comparable performance with far lower training and inference expenses. Experiments at Du Xiaoman show that a 13‑billion‑parameter model fine‑tuned for finance can outperform a generic 70‑billion‑parameter model on certain financial tasks.