On November 2, at the opening ceremony of Hong Kong FinTech Week 2026, Du Xiaoman, together with the Hong Kong University of Science and Technology and the Hong Kong Generative AI Research and Development Centre (HKGAI), jointly released Xuanyuan 4.5, a finance-enhanced large model. The publisher says it is the first finance model to apply self-evolving RSI technology to model training and to adapting agent scenarios to goals, and that it comprises both a large language model and a multimodal model. Competition among finance models is shifting from general capability toward who understands a specific business better, and who can keep evolving alongside that business.
At the core of Xuanyuan 4.5 is an in-house system called Xuanyuan-Machine, which takes part in agent data synthesis, supervised fine-tuning, and reinforcement learning, letting the model keep adapting to finance tasks. The logic is easy to grasp: the definitions, rules, and customer phrasing in finance tasks change often, and a model trained only once will quickly fall behind; only by having machines take part in the data and training loop can the model keep adjusting inside the business. The partnership has continuity: in March 2025, Du Xiaoman and the Hong Kong University of Science and Technology set up a joint laboratory and worked with HKGAI on applying large models to credit risk control.
Customer acquisition, risk control, and customer management: how to read the effectiveness claims
Du Xiaoman says Xuanyuan 4.5 improves results by 30% in core finance scenarios such as customer acquisition, risk control, and customer management, at only about 10% of the inference cost of top general models. The release also says more than 1,000 agents and more than 5,000 skills have been built on Xuanyuan, with daily token usage exceeding 100 billion. These numbers should be discounted: the coverage discloses no evaluation methodology, no list of the general models used for comparison, and no reproducible test details, and the "first" claim is likewise the publisher's own description, not yet independently verified.
For financial institutions, there is a more practical hurdle beyond the effectiveness numbers: figures must not be wrong, definitions must be consistent, every step must be traceable, and model outputs must pass compliance and audit checks. Whether Xuanyuan 4.5 can work in real business in Hong Kong depends on how it performs under these hard constraints, not only on the effectiveness claims made at the launch. What is worth watching next is whether concrete Hong Kong deployment cases and evaluation details are made public, and whether the cost advantage holds up under real workloads.
