Zhang Lei, Chief AI Scientist and Senior Vice President of Yixin: Building a "Chinese Solution" for Auto Finance with Full-Stack AI Capabilities | 36Kr 2025 AI Partner Conference for All Industries

On August 27, the 2025 AI Partner Conference for All Industries, jointly hosted by 36Kr and the China Europe International Business School, grandly kicked off at the Zhongguancun Software Park in Beijing. The theme of this conference is "Chinese Solutions", which is divided into two major chapters: "Chinese Solutions" and "Who Will Define the Next AI Era". Centering around four major topics, namely "The Golden Age of Chinese Innovation", "Can Super Intelligent Agents Become the Core Form of the Next Generation of AI?", "Chinese Solutions Reshape the Global Technological Competition Landscape", and "The Prosperous Scene of the Integrated Innovation of AI and All Industries", the conference comprehensively presented the latest breakthroughs and ecological systems of Chinese AI, shared the growth path and future prospects of Chinese-style AI, and explored the innovation models of Chinese solutions.

On that day, Zhang Lei, the Chief AI Scientist and Senior Vice President of Yixin, gave a keynote speech titled "AI Reshapes the Automobile Financial Service Ecosystem".

The following is the content of the speech, sorted and edited by 36Kr:

Dear guests and teachers:

Good morning! I am very honored to be here to share with you Yixin's practices and insights in the field of "AI Reshapes the Automobile Financial Service Ecosystem".

First, let me briefly introduce Yixin. We are an AI-driven fintech platform and a leading listed automobile fintech company in China. Our annual transaction volume is approximately 70 billion yuan, and we have cumulatively invested over 2 billion yuan in R & D and AI fields. In terms of technical strength, our large model is the only one in the industry that has passed the national certification and filing. We are also the first company to achieve the local deployment, full application, and model open - source of DeepSeek. We always believe that to reshape the industry and service ecosystem, a strong AI foundation is essential, and AI capabilities are the core.

Looking back at the development history of AI, we can divide it into different stages:

In the early days of 1.0 discriminative AI, all business judgments were based on rules. For example, the expert scoring cards commonly used in the financial scenario, and the early days of search, advertising, and recommendation (such as the early days of Taobao) also mostly relied on rules, and the business process was static. In the later stage of discriminative AI, although small models such as traditional machine learning were introduced, the process remained static.

Entering the 2.0 deep - level AI stage, large models began to participate in business judgments, covering aspects such as feature analysis and modeling cooperation. However, the process still did not break through the static limitation. It is not difficult to find that in the human - machine cooperation models in the 1.0 and 2.0 stages, it has always been "humans take the lead, and AI plays a supplementary role", and AI mainly has a profound impact on people's business workflows.

Based on our practices, we believe that for enterprises to deeply apply AI, they need to have full - stack AI capabilities. That is, they need to have AI infrastructure (including computing hardware and integrated training and inference software) as well as AI application capabilities. Relying on these resources, pre - trained and post - trained models can be built. Then, domain models of different sizes can be generated based on the foundation, and platform products can be created. These products are like "building blocks", and front - end business applications only need to combine them flexibly.

The industry - specific model matrix we have built covers pre - trained and post - trained models from the bottom up, as well as domain models of four sizes. The reason for dividing into different sizes is that domain applications have dual requirements for effectiveness and efficiency. For example, in high - real - time scenarios, an overly large model size will affect the response efficiency. In addition, we also have a 7B acoustic model (used for high - fidelity speech synthesis and cloning), two inference models (focusing on logical reasoning), a multi - modal model, and an Agentic Model that supports multi - intelligent agent cooperation.

Why do we need vertical domain models? There is a "trade - off" between general models and vertical domain models. Although general models have a wide coverage, they consume a lot of resources and have a shallow understanding of vertical domains. Vertical domain models can accurately understand the needs of scenarios. For example, if a user asks, "I've driven a gasoline - powered car for ten years and want to switch to an electric car, but I'm worried about the battery safety of unknown manufacturers", general models usually answer around the comparison between gasoline and electric cars, while our automobile finance vertical domain model can accurately capture the user's core concern of "fear of fire" and explain the risk points in a targeted manner, reflecting a deep understanding of the industry.

In March this year, based on the 72B - sized model and our self - developed training methodology, we open - sourced an inference model with domain knowledge. Its mathematical and logical abilities are comparable to the top - notch general models at that time. We also provided two versions, full - size and quantized - size, to help the industry jointly build AI capabilities.

Our core goal is to build a "full - link AI decision - making engine" and serve the business in a product - oriented way from beginning to end. In 2018, we released a decision - flow platform that supports traditional machine learning models such as LR and XGBoost to empower risk control. In 2019, we launched a model platform to adapt to the needs of large - scale deep learning. In 2020, we launched a robot platform to realize the productization of algorithms and models. In 2022, to adapt to the development of AI, we overthrew the third - party procurement system and self - developed an AI - native call center system. Since 2023, we have successively released a 7B small - sized large model, a text - to - text and multi - modal large model that has passed the national filing, and fully applied the large model to the upgrade of robot products in 2024. We also launched a new media creation platform covering marketing scenarios such as short videos and live broadcasts.

The essence of finance is "the commercialization of risks". We have broken down the automobile finance business into three major links and empowered them deeply with AI:

Before financing (channels + application intake): Use AI search capabilities to automatically generate channel analysis reports to assist decision - making. When taking in applications, use multi - modal models to automatically retrieve and extract user information for automated entry.

During financing (risk control + intelligent link): In addition to personalized risk models, use the robot platform to cover all aspects of risk control. We have uniquely created an "intelligent link" model and algorithm to match users with the optimal financial solutions. We also launched "end - to - end risk control", directly inputting raw information such as text, pictures, audio, and video into the model, allowing the model to extract features independently. At the same time, we integrate traditional interpretable models to avoid information loss caused by manual feature extraction.

After financing (customer service + asset management): On the customer service side, in addition to automatically answering phone and IM consultations, we also analyze users' voiceprints and intonations to predict the risk of complaints and dispatch orders in advance. On the asset management side, we use AI to formulate personalized strategies and match corresponding solutions for different risk levels.

The automobile finance industry has long faced the problem of "complex decision - making + low efficiency of long - link workflows". We believe that only by entering the 3.0 Agent AI stage can we break through. In this stage, large AI models are responsible for making judgments, and the business process changes from static to dynamic. AI evolves in real - time through interaction with customers and the environment to adapt to the optimal process. The human - machine cooperation also shifts from "human - centered" to "machine - centered", and the business process is designed around the machine.

Currently, we are promoting the agentization in a product - oriented way, with the core being "AI Agent Business Interface + AI Risk Control Intelligent Chain". Taking the pre - approval process of application intake as an example, we have abstracted three major intelligent agents: "Outbound Call Assistant, IM Assistant, and Approval Assistant". The Outbound Call Assistant automatically contacts users to confirm their intentions and qualifications. At the same time, it synchronously links the IM Assistant to add users as friends and follow up on the collection of materials (text, images, videos, etc.) throughout the process. After the materials meet the standards, the IM Assistant and the Approval Assistant cooperate to complete the approval. If additional materials are needed, the IM Assistant will communicate to obtain them. The whole process does not require manual intervention and is completed through the cooperation and arrangement of intelligent agents. Although the process is simple, a large number of technical and algorithmic difficulties have been overcome behind it.

Now, Yixin's fintech solutions have achieved "domestic application + global expansion". Globally, China and the United States lead in AI strength, and our automobile finance AI capabilities have been verified by the market and are highly competitive in the global industry.

That's all for my sharing. Thank you!