As digital transformation enters deeper waters, holistic intelligence has become an industry-wide consensus. As a pioneer, China Merchants Bank has taken the lead in independently developing a full-stack large model technology system, offering the industry a practical paradigm and reusable methodology for evolving from “point intelligence” to “system intelligence.”
The bank has built a fully self-controlled large model stack spanning infrastructure, models, and applications, achieving tangible progress across each layer. At the infrastructure level, a series of foundational innovations have driven industry-leading performance in core computing utilization and single-card token throughput.
Built entirely through in-house research and development, the platform delivers leadership in functionality, performance, and cost efficiency. It features a financial-grade heterogeneous computing cloud foundation, a low-latency, high-efficiency inference platform based on heterogeneous cards, and a cluster-based training platform supporting agentic reinforcement learning.
Proprietary training and inference frameworks have improved end-to-end inference performance by more than 50 percent. The bank has also contributed 40 key features to major open-source projects and earned maintainer status in two leading open-source communities, underscoring its role in advancing the broader technology ecosystem. From a cost perspective, token processing expenses are approximately 70 percent of those charged by mainstream public cloud providers.
On the model layer, China Merchants Bank has developed a structured model matrix that balances quality, efficiency, and security. It has deployed more than 40 cutting-edge open-source foundation models across multiple modalities and parameter scales, optimizing operators to enhance inference performance and resolving heterogeneous computing adaptation challenges for models such as DeepSeek and Qwen.
At the same time, the bank has extensively reengineered model architectures to create more than 60 specialized domain models covering customer service, client management, middle- and back-office operations, and research and development. These tailored models significantly improve accuracy and operational performance in financial scenarios.
At the application level, the bank has established one of the most comprehensive and deeply integrated business-technology collaboration systems in the industry. Through planning optimization, token consumption has been reduced by 55 percent; context compression has tripled the number of conversational turns; and parallel tool execution has shortened processing time by 13 percent.
More than 12,000 users are covered by the platform, with business staff accounting for over 40 percent, reflecting a broad-based adoption beyond technical teams. Supported by a data-model-evaluation toolchain, application development cycles have been compressed to as little as eight days, enabling rapid iteration and value realization.
This full-stack capability has translated technological strength into frontline business value. To date, more than 800 application scenarios have been deployed across retail banking, corporate banking, risk management, operations, office administration, and R&D. These applications span knowledge Q&A, report processing, risk and compliance review, document verification, and software development, comprehensively improving employee productivity, lowering operational thresholds, and enhancing customer experience.
In software development, the bank has launched its proprietary DevAgent, an intelligent R&D agent built on a multi-round “perception–planning–execution–feedback–evolution” ReAct framework. By understanding natural language instructions, sensing the developer’s coding environment, and retrieving enterprise knowledge, DevAgent delivers task-level development capabilities, including cross-file and large code block generation. It now completes tens of thousands of development tasks each month, significantly accelerating product iteration and improving engineering efficiency.
In retail banking, an AI-powered investment research assistant provides relationship managers with quantitative analysis and intelligent product screening. Tasks that once required analyzing more than 1,000 indicators and hours of manual report consolidation can now be completed within minutes, dramatically improving responsiveness and professionalism in client service. In wholesale banking, a digital assistant supports data queries, analytics, and list retrieval for branch and head office staff, reducing high-frequency data retrieval time from minutes to seconds.
Beyond internal efficiency gains, the bank has embedded AI deeply into customer-facing services. Upholding its commitment to inclusive finance, China Merchants Bank integrates technological innovation throughout the service lifecycle. Its mobile app offers real-time multilingual translation, enabling foreign residents in China to switch key interfaces and product information instantly into different languages.
For Chinese enterprises expanding overseas, AI-powered document processing and risk control models have compressed traditional account opening and due diligence timelines to roughly one-third of their previous duration, enhancing both efficiency and cross-border risk management.
In inclusive finance, the bank has introduced voice-enabled services that allow customers to complete transactions with a single spoken instruction. Multimodal technology converts complex wealth management graphics into audio prompts, enabling visually impaired users to independently conduct financial operations. An AI-powered telephone assistant now supports more than 300 business scenarios through voice interaction, improving accessibility for elderly customers and dialect speakers.
China Merchants Bank’s experience validates the feasibility of diverse technological pathways and provides the financial industry with a mature, reliable, and scalable engineering methodology for large model implementation.
Looking ahead, the bank will continue to anchor its strategy in technological innovation, expanding both the breadth and depth of its services. On a foundation of security and compliance, it aims to build more adaptive, human-centered, and sustainable financial experiences, growing together with customers, partners, and society in an increasingly intelligent era.
Source: the economic observer, sz gov cn, cmb china, eeo
