Ant Financial's Large Financial Model Released: From Ant to the Future of Fintech Revolution

[baoshi.rao]

On September 8, Ant Group officially released its large financial model. It is understood that the Ant Financial Large Model is based on Ant's self-developed foundational model, deeply customized for the financial industry, with an underlying computing power cluster reaching tens of thousands of GPUs, focusing on real financial scenario needs.

Wang Xiaohang, Vice President of Ant Group and head of the financial large model, told TechWeb and others that large models are bringing transformative experiences to the financial industry. "Every key function in the financial business chain is worth redoing with large model technology."

Currently, the Ant Financial Large Model has been fully tested on Ant Group's wealth and insurance platforms. Two products based on the financial large model's capabilities are also in internal testing and will be launched after completing relevant regulatory filings.

"In a rigorous field like finance, there are very few industrial-grade products. What we are releasing today is an industrial large model, hoping to solve industry problems and become a practical and professional force in this round of technological competition," said Wang Xiaohang.

Focus on Practicality

In June this year, it was reported that Ant Group's R&D team was developing its own language and multimodal large models. At the time, Ant responded that "the information is accurate." Now, Ant's first public step with its large model is directed at finance, aiming for large-scale industrial applications, with a focus on practicality.

According to He Zhengyu, Chief Technology Officer of Ant Group and President of the Platform Technology Business Group, as early as September last year, during an internal strategic meeting, the team decided that AI must be upgraded around large models. Moreover, the team believed that large models must be combined with industrial applications.

With years of deep experience in the financial industry, Ant has unique advantages in developing financial large models. In terms of computing power, the Ant Financial Large Model is based on Ant's foundational model, deeply customized for the financial industry. It is reported that Ant's foundational model platform features a heterogeneous cluster of tens of thousands of GPUs.

In terms of knowledge, the Ant Financial Large Model is built on trillions of tokens of general data, infused with hundreds of billions of tokens of financial knowledge, and extracts over 600,000 high-quality instruction data points from more than 300 real-world industry scenarios, forming a superior data asset for financial-specific task performance optimization.

Additionally, the Ant Financial Large Model can accurately call various digital financial tools within Ant's ecosystem by understanding user language. On the wealth management side, it includes six major services such as product selection, product evaluation, market analysis, and asset allocation. On the insurance side, it covers more than ten intelligent services, including product interpretation, family planning, intelligent underwriting, and intelligent claims.

According to Wang Xiaohang, based on extensive practice in financial scenarios, the Ant Financial Large Model has developed an architecture driven by "large model + knowledge + service," which is already being tested internally for financial intelligence scenarios.

How to Prevent Large Language Models from 'Hallucinating'

Due to biases and misleading information in training data, large language models may produce inaccurate or unreasonable responses in human-computer interactions, a phenomenon often referred to as 'model hallucination.'

However, in a highly regulated field like finance, financial large models must ensure the rigor of domain knowledge, professional logic, and controllability of content generation to deliver real industry value. This is the core challenge financial large models must address.

Wang Xiaohang, in an interview, acknowledged that finance demands high levels of expertise, logical rigor, and compliance. Native large models still fall short of the financial industry's requirements, and there is no foolproof method to prevent hallucinations.

"To address this, we adopted a dual-driven approach combining structured data from knowledge graphs with parametric data from large models. By leveraging extensive financial knowledge graphs, we aim to ensure the model's professionalism and rigor," Wang explained. This involves knowledge injection, consistency alignment, and post-generation verification, forming a systematic process.

Wang believes the financial industry is not yet mature enough to fully harness the opportunities offered by large models. "This requires fintech companies, including internal tech teams within financial systems, to transform technology into products and platforms. I estimate this will take about 1-2 more years."

Pilot Products Under Internal Testing

"We hope large models can bring transformative experiences to the financial industry," Wang stated. Zhi Xiaobao 2.0 and Zhi Xiaozhu are applications developed by Ant Group based on financial large models, currently in internal testing.

Zhi Xiaobao, an intelligent wealth management AI codenamed 'Anna,' was initiated in 2018, the same year GPT-1.0, the earliest version of ChatGPT, began development. Now integrated with large models, its financial knowledge database has reached billions of data points.

In a live demo, when a user asked, 'Which industries will AI impact given its current popularity?' Zhi Xiaobao 2.0 synthesized insights from millions of reports and authoritative sources to provide an answer. For queries like 'Should I buy now?' it utilized tools like portfolio diagnostics and allocation analysis to assess the AI sector's market conditions, advising the user to reduce positions and avoid irrational trading behaviors.

When the user followed up with 'Still don’t understand,' Zhi Xiaobao 2.0 rephrased its explanation, first defining the term 'satellite fund' and then clarifying the importance of diversification to balance risk and returns.

Beyond being smarter, Lu Xin, the technical lead of Zhixiaobao, explained that with enhanced language capabilities, Zhixiaobao 2.0 has also improved its emotional intelligence. Its financial intent recognition accuracy has reached 95%, enabling it to understand user emotions, engage in high-quality multi-turn conversations, and even adjust its communication style to be more professional or colloquial. After answering user queries, Zhixiaobao, powered by a large model, automatically 'reflects' on the accuracy of its responses and self-corrects.

Previous data shows that as of 2022, Zhixiaobao 1.0 (a retrieval-based AI) has served over 300 million users and answered more than 1.7 billion financial questions.

Unlike Zhixiaobao, which targets individual users, Zhixiaozhu is an intelligent business assistant designed for B2B financial industry experts. It provides in-depth intelligent services in areas such as investment research analysis, information extraction, professional content creation, business opportunity insights, and financial tool usage.

For example, test data shows that 'Investment Research Zhixiaozhu' can assist each research analyst in extracting financial logic and insights from over 100 reports and news articles daily, as well as reasoning and attributing more than 40 financial events, significantly boosting analysis efficiency. Additionally, Zhixiaozhu can largely replace basic financial engineering coding, greatly enhancing quantitative research efficiency. With the help of 'Service Zhixiaozhu,' financial advisors and insurance agents can expand their effective client management radius by over 70% per person.

It was revealed that Zhixiaobao 2.0 has been in internal testing for nearly six months and will be launched after completing the filing process. Zhixiaozhu is currently undergoing co-development testing with Ant Group's partner institutions and will be officially opened to them once mature.

Commercialization Still Requires Time

At the Bund Summit, Ant Group disclosed its '1+1+2' financial AI matrix for the financial industry: one financial large model, one Fin-Eval financial AI task evaluation set, the personal financial assistant Zhixiaobao 2.0, and the expert business assistant Zhixiaozhu.

Fin-Eval represents real-world industry demands and is currently the most comprehensive and professional evaluation set in the financial AI field. According to Wang Xiaohang, Ant Group has officially opened Fin-Eval to the public, hoping to promote collective technological progress in the industry.

'Our approach to technology openness has always been to open up when a technology matures,' Wang Xiaohang explained. 'When the productivity applications of Ant's financial large model become more mature, we will promote it to the industry on a larger scale and broader scope.'

When discussing the future business model of Ant's financial large model, He Zhengyu stated in an interview that AI large models are still in their early stages, making it difficult to discuss commercialization definitively. On one hand, the technology is still in its 'infancy,' and on the other, it requires more time.

He Zhenyu believes that product and software monetization represents a business model. Commercial attempts by AIGC products like Miaoya Camera, as well as software-delivered products such as Word, may provide valuable references for large model commercialization. The current biggest challenge lies in accurately measuring and providing feedback on the value created by large models, which requires careful consideration.

It is reported that Ant Group will continue to explore and enhance five key capability directions for large models:

1. Building a high-quality data annotation team to establish a premium data system
2. Advancing fundamental large model algorithms and developing efficient, green engineering capabilities to improve model functions like logical reasoning
3. Evolving from general language models to universal multimodal models, progressing from basic common knowledge to comprehensive expertise
4. Establishing efficient evaluation standards and systems for large models to accelerate iteration speed
5. Developing robust security capabilities for large models to ensure their healthy and sustainable development.