NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Du Xiaoman Takes a Finance Model to Hong Kong, Claiming One-Tenth the Inference Cost

Du Xiaoman, HKUST and HKGAI launched Xuanyuan 4.5 at Hong Kong FinTech Week, a finance model the partners say cuts inference cost to a tenth of a general model's.

2026-11-21 · 1011 words · NeuroAI
Du Xiaoman Takes a Finance Model to Hong Kong, Claiming One-Tenth the Inference Cost

A loan officer used to spend ten minutes checking a file before giving an answer. A new system says it can cut that wait to two minutes and keep working after the branch closes. The companies behind it chose an unusual place to make the announcement: the opening ceremony of a fintech week in Hong Kong.

On 2 November 2026, at the 2026 Hong Kong FinTech Week, Du Xiaoman (度小满), the Hong Kong University of Science and Technology and the Hong Kong Generative AI Research and Development Center (HKGAI) jointly launched Xuanyuan 4.5 (轩辕4.5), a financial enhanced large model (大模型). It is a corporate move with a product attached: a Chinese fintech using a university partnership to sell AI into a market that has its own rules about credit and data.

What was launched

Xuanyuan 4.5 is described by its makers as a financial enhanced large model with two parts, a large language model and a multimodal model. The technology they highlight is called RSI, which the partners say lets the model keep adapting to financial tasks rather than staying fixed after training.

The companies put a name on the internal machinery. A self-developed component called Xuanyuan-Machine takes part in agent data synthesis, supervised fine-tuning and reinforcement learning. For problems that need several steps of analysis, the system combines long-chain reinforcement learning, rewards based on standard credit workflows and online policy distillation. The partners say this was built on more than 100 billion tokens of scenario data inside a compliant financial agent (智能体) environment.

The launch continues a partnership that predates the product. In March 2025, Du Xiaoman and HKUST set up a joint artificial intelligence laboratory, working with HKGAI on model research and on how models might be used in credit risk control. Xuanyuan 4.5 is the first visible output of that arrangement to reach a public stage.

The claim, and how to read it

The headline numbers are the partners' own. Securities Times (证券时报), reporting from the launch, records these claims:

  • Xuanyuan 4.5 improves results by 30 percent in core financial scenarios including customer acquisition, risk control and customer management.
  • Its inference cost is about one tenth that of a top general-purpose model in the financial scenarios tested.
  • It is the first self-evolving financial enhanced large model in the industry — a "first" asserted by the speakers at the event, not by an independent body.

None of those figures comes with a public benchmark, a test set or a methodology that an outside party could reproduce. The comparison is drawn against "general top models" that are not named in the announcement. Until a third party runs the same tasks on the same data, the honest description is that the companies claim a 30 percent improvement and a tenth of the cost.

Du Xiaoman also gave an account of its broader deployment. The company says it has built more than 1,000 agents and over 5,000 skills, with daily token use above 100 billion. It says risk-strategy iteration cycles have been cut by more than half, short-term overdue risk has fallen by about 50 percent and overall annualised overdue risk by 12 percent. It says an AI credit reviewer outperforms 80 percent of human reviewers, cutting a customer's wait from ten minutes to two; that AI post-loan specialists achieve a collection rate more than 8 percent above conventional bots; and that agents now handle 80 percent of high-frequency customer-service questions on their own. These are internal operating figures disclosed by the company, with no audit attached.

Why Hong Kong matters

The venue is part of the story. Du Xiaoman calls Hong Kong its international headquarters, and its chief executive, Zhu Guang (朱光), used the launch to say the company would deepen research cooperation with HKUST and work with HKGAI to take financial enhanced models into overseas markets. Guo Yike (郭毅可), HKUST's chief vice president, framed the project as a way to bring academic AI research into real industrial settings.

That makes this a corporate strategy story rather than a pure model release. A Chinese fintech is using a Hong Kong university partnership, a local research centre and an international fintech event to position a product for cross-border sale. Whether overseas regulators will accept a credit model trained on Chinese lending data is a separate question, and the announcement did not address it.

According to Li An, chief scientist at BrainNet (脑机网), the useful test for a vertical model is not its score on a vendor's own task list but whether the cost advantage survives when the model meets a different regulatory and customer environment. For Xuanyuan 4.5, the Hong Kong debut is the start of that test, not the result.

Honest limitations

This account is based on reporting by Securities Times (证券时报) from the 2 November launch and on statements attributed there to Du Xiaoman and HKUST executives. It is not based on a technical paper, a benchmark report or audited company data. The 30 percent improvement, the one-tenth cost figure, the "first" designation and all operating metrics are company claims reported at an event; none has been independently verified, and no test methodology or baseline model is named. The exact scope of Xuanyuan 4.5's availability, its price and any customer contracts were not disclosed. This article does not assess whether the model complies with financial regulations in Hong Kong or in any other jurisdiction, and it makes no judgment about the credit practices of the companies involved.

What readers can do now

  • If you are evaluating a vendor's vertical AI model, ask which tasks were tested, which baseline was used and who ran the test. A percentage with no method behind it is a marketing figure.
  • If you work in lending or customer service, separate "the agent handles 80 percent of queries" from "the agent handles 80 percent of queries correctly." Ask for the error rate on the hard cases.
  • For anyone watching Chinese AI companies going abroad, follow the regulatory filings and named pilots, not the launch-stage cost claims. Cross-border finance is where a model's origin data matters most.

Related coverage

Cite this article

NeuroAI (脑机网), "Du Xiaoman Takes a Finance Model to Hong Kong, Claiming One-Tenth the Inference Cost," 21 November 2026, https://neuroai.site/a/na-corp-duxiaoman-hongkong-finance-model

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Companies & Stack” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories