In recent discussions among banking professionals, the rising trend of DeepSeek was the hot topic on everyone's lipsDespite not yet reaching mainstream application levels, many are considering how to harness the wave of innovation it represents.

DeepSeek has rapidly gained popularity worldwideNotably, Jiangsu Bank has already adopted the new model, and several other financial institutions are utilizing it in their marketing effortsThe breakthrough technology offered by this new tool in the realm of banking models is prompting industry insiders to reflect deeply.

Experts point out that as DeepSeek and other AI models continue to evolve, the cost of single inference calculations will decrease while computational power infrastructure will gradually improve

Advertisements

This progression signals a pivotal shift from "technical feasibility validation" to "economic viability realization" in applications of large-scale models.

Banks Dive into the DeepSeek Experience

At the beginning of 2025, DeepSeek-R1, developed by the AI firm Deep Seek under Hangzhou-based Quantum Solutions, debuted with much fanfareIt quickly became a viral sensation over the Spring Festival periodUsers can activate the deep thinking and networking options with a simple command, connecting them with a "cyber friend" possessing a wealth of knowledge at the other end of the lineIts localized database endows it with superior language generation capabilities and a significantly lower price point compared to ChatGPT, leading to its explosive growth seemingly overnight.

Capitalizing on the buzz around DeepSeek, some banks have launched unique promotional campaigns

Advertisements

For instance, Haian Rural Commercial Bank recently published a feature where it introduced itself in an engaging dialogue with DeepSeek, covering dimensions like capital strength and market share, effectively capturing attention.

Beyond entertaining the public, DeepSeek has sparked deeper contemplation among industry expertsIn interviews, many highlighted the technological breakthroughs that the DeepSeek model represents as a significant focal point.

In recent years, the consensus among the banking industry has been the importance of advancing digital finance and embracing digital transformationSeveral listed banks have expressed positive sentiments regarding the advancement of large AI models in their recent financial reports, revealing planned initiatives and progress in developing these technologies.

In this landscape, Jiangsu Bank currently holds the distinction of being the first publicly listed bank to adopt DeepSeek

Advertisements

According to disclosures from the "Su Yin Digital Finance" public account, Jiangsu Bank successfully localized and fine-tuned the DeepSeek-VL2 multimodal model and the lightweight DeepSeek-R1 inference model, deploying them in scenarios for intelligent contract quality inspection and automated valuation reconciliation.

"DeepSeek's diverse capabilities, coupled with its strong logical reasoning and natural language processing abilities, allow for cost-effective deployments," pointed out Su Xiaorui, a senior researcher at Su Xi Zhi Yan.

According to research from Zheshang Securities, the entire training process for DeepSeek-V3 required only around 2.8 million GPU hours, whereas Meta's Llama3-405B took approximately 30.8 million GPU hours for training

• How the U.S. Can Lower Treasury Yields
• How to Lower U.S. Treasury Yields
• Revitalizing the Auto Market through Trade-Ins
• Nissan and Honda's Two-Month Split
• Honda-Nissan Merger Plan Falls Through

The training cost for DeepSeek-V3 is about $5.576 million, in stark contrast to the hundreds of millions spent by OpenAI on their ChatGPT language model, GPT-4.

What accounts for the widespread attention from the financial sector towards this newly released model? Beyond its affordability, Su Xiaorui emphasized that DeepSeek's parent company has a strong background in private equity, granting it a more robust financial DNA compared to other AI models.

Meng Can, the chief analyst for computers at Guojin Securities, noted that DeepSeek's core advantages lie in its customer-centric approach and understanding of contextFrom a business perspective, the AI's reasoning path combined with human oversight increases the rigor and professionalism of banking decisions.

The Importance of Safety and Accuracy

A report released earlier by the China Banking Association highlighted the natural synergy between finance and artificial intelligence, asserting that AI's large model technology could greatly leverage the banking industry's vast datasets, which are rich in applicable scenarios for AI model technologies

Currently, AI large models are catalyzing a comprehensive transformation in banking services, marketing, and products, accelerating the arrival of the "future bank."

As generative AI technology accelerates its penetration into the financial sector, banks have advanced from initial tech validation to the large-scale application stage of model explorationWhat, then, are the paths for the implementation of these AI large models in banking? What risks and challenges are ahead?

"Since the advent of ChatGPT, banks have allocated significant resources towards computational power reserves and the development of proprietary financial large modelsAs DeepSeek and future AI models evolve, the decreasing cost of single inference calculations signifies an increase in capacity of computational power infrastructure, marking a critical shift from 'technical feasibility validation' to 'economic viability realization.'" A veteran bank expert relayed to reporters.

Managing risks associated with the introduction of large AI models is crucial since banks are institutions that deal with risk management.

"In the current applications of AI, we face notable challenges such as AI hallucination — the phenomenon where AI misinterprets input data, generalizes excessively, or makes inaccuracies during reasoning

Essentially, AI sometimes imagines non-existent elements or information, akin to how humans perceive things in dreams or hallucinations," the aforementioned banking expert elaborated"This necessitates the identification of erroneous information; we can utilize DeepSeek's open-source explainable toolchain to enhance transparency in the decision-making pathway of the model and isolate the organizational knowledge base to prevent polluting information, averting a vicious cycle."

Meng Can also highlighted that while employing large models, banks must prioritize data security, particularly regarding boundaries surrounding sensitive data"Every interaction between the AI large model and clients involves client privacyAdditionally, passive data collection may raise intellectual property concerns," he cautioned.

"In the long run, banks may gradually complete the transformation of the financial services ecosystem based on AI applications," Meng Can suggested

"From a productivity viewpoint, the question-and-answer based model will transition to comprehensive service delivery, enhancing service efficiency, accuracy, and customer satisfaction while lowering decision-making risksFrom the perspective of production relations, banks' organizational structures may undergo changes, resulting in more harmonious relationships with clientsFurthermore, the adaptation of software and hardware in banking services will evolve."

In interviews, Su Xiaorui expressed that the application of large AI models holds substantial long-term potential across various specific scenarios such as intelligent marketing and risk managementIt is anticipated that more licensed institutions will join the AI upgrade wave, improving the quality and efficacy of traditional financial operations while ensuring financial security and user account safety.

Why are there currently few banks employing the DeepSeek tool? Industry insiders have pointed out that many banks developed financial large models quite early on

Although they implemented a layer-coupled architecture, adjustments to interfaces and retraining of knowledge bases can hinder model transitionsHowever, the open-source technology framework has truly ignited an AI arms race across the industry, and as more cloud computing platforms support DeepSeek, the migration of models will accelerate.

"The core competitiveness for banks developing large models stems from three key aspects," the insider concluded"Firstly, can the organization truly form a comprehensive knowledge base that connects knowledge silos, fragmented scenarios, and feedback gaps to create a vector knowledge repository from business operations and customer feedback? Secondly, can they establish robust firewalls for data privacy protection and risk isolation, interacting with data, models, and applications to build comprehensive lifecycle management for the models? Thirdly, the capability for ecosystem collaboration is crucial; they must base their services on ample feedback from business scenarios and present information simply and effectively to resolve clients' challenges."