Application of large models in banks: "Small is beautiful, multiple models for one enterprise, practicality first"

The previous article introduced the overview of big model applications of leading banks. In their annual reports, each bank discussed their innovations from multiple dimensions, including the overview of big model introduction, application scenarios, R&D efficiency improvement, and key performance data. They were excellent and distinctive, and were impressive.

Since small and medium-sized financial institutions do not have the resources and teams of leading banks, it is difficult for them to directly copy the practices of these large banks in the path of enterprise-level AI/ big models. Recently, IBM , a long-established technology supplier in the financial industry, proposed the view that " small is beautiful, one enterprise has multiple models, and practicality is the priority " for enterprise-level AI . This may provide another perspective for small and medium-sized financial institutions.

Last week, at the Think 2025 conference held in Boston , IBM put forward the core view of enterprise AI : small is beautiful , and small refers to domain-specialized small models. " One enterprise, multiple models " , in the future, not only one large model will be deployed, but multiple small models will work together in different scenarios (such as one for marketing, one for customer service, and another for IT  operations). " Practicality first " does not blindly pursue the largest model size and the most parameters, but pursues the best fit between performance, efficiency and business scenarios.

IBM proposed the enterprise AI  strategy of " specialized small models first " at the Think 2025 conference . IBM  CEO Arvind Krishna  has repeatedly emphasized that enterprise-level AI  applications require models customized for enterprise data and business scenarios , rather than simply pursuing the largest general-purpose models.

IBM points out that less than 1%  of enterprise data is currently used by large general-purpose models . “ To unlock the value of the remaining 99% of enterprise data, AI  methods tailored to enterprise scenarios are needed _[1] ” . At the same time, smaller, open, and dedicated AI models  can achieve ROI faster .

Why choose a small model?

IBM provides multiple reasons for choosing a small model, mainly including:

·Cost and efficiency : Small models require far fewer computing resources than super-large models, and their operating costs are greatly reduced. For example, IBM  revealed that its dedicated small models can be more than 30  times cheaper than large models with hundreds of billions of parameters when inferencing. Small models only contain data related to specific tasks, thus saving computing and energy consumption.

·Deployment and Performance : Small models are smaller, faster, and easier to deploy in a variety of environments. They can run flexibly on local servers, edge devices, or the cloud, without relying on large-scale computing power. Small models are not only " much faster and more cost -effective " , but can also be executed anywhere on demand. In scenarios with limited computing power (such as edge devices such as smartphones and industrial sensors), small models can process data on-site and instantly release the value of edge data.

·For enterprise data and scenarios : Large general models are often " incapable " of being used in enterprise segmentsand have limited generalization capabilities. IBM  emphasizes the use of small models customized for specific areas, which can learn more deeply about the enterprise's own data, thereby improving accuracy and relevance.

·Controllability and security : Small models are easier to " see out of the box " , which is convenient for enterprise auditing and tuning, and reduces black box risks. IBM  pointed out that many general models are closed-source black boxes, and enterprises cannot verify the internal mechanisms. Small models are mostly deployed in an open source and transparent manner, which helps enterprises retain data sovereignty and intellectual property rights. Small models can also be customized and trained according to corporate compliance and privacy requirements, which improves the credibility of results and reduces the risk of bias and vulnerabilities.

·Specialized performance advantages : Small models for specific tasks can sometimes outperform large general-purpose models in terms of accuracy. IBM  found in internal testing that domain models with 3 billion to 20 billion parameters, after targeted tuning, can match or even outperform large models with hundreds of billions of parameters in terms of business accuracy [2] _. At the same time, because the models are smaller and more specialized, they also have obvious advantages in terms of inference speed, response latency, and other aspects.

Comparison with the dominant trend of large models

The AI ​​industry was once enthusiastic about super-large-scale general models. IBM 's point of view is not to completely deny large models, but to emphasize the complementarity of the two. IBM CEO Arvind Krishna  made it clear that small models are not to replace large models, but to be used " in combination with large models " , especially in enterprise scenarios to optimize according to local conditions. IBM  believes that large models are useful for some complex tasks, but for most daily enterprise needs such as classification and summarization, super-large-scale models are not needed. IBM  even cited research showing that the scarcity and cost of training data have made it uneconomical to continue to expand blindly. In the future, AI  applications will rely more on " where the data exists " rather than how big the model is.

In contrast, industry trends have also begun to shift from " bigger and better " to more pragmatic exploration. Gartner predicts that by 2027 , enterprises will use customized small models three times more often than general-purpose LLMs  [3] . Forrester  also predicts that the deployment of small models will surge by more than 60% this year , especially those trained for specific business terms and processes. Mainstream technology media and institutions have paid attention to the rise of small models: for example, MIT Technology Review  listed small language models as one of the top ten technological breakthroughs in 2025  , calling them " faster, cheaper, and more efficient [4 ] .

The idea of ​​giving priority to small models has caused a great response in the industry

The industry has generally responded positively to IBM’s  small model strategy, with most experts and companies believing that this is the direction of development. TheCUBE  , a technology analysis agency , pointed out that IBM  ’s 2025 industry small model strategy validates the “Gen AI  power law hypothesis ” they proposed in 2023  : the long tail of future generative AI  applications will be countless small models for different data locations [5] .

Technology media WIRED reported that not only IBM , but also Microsoft, Google, OpenAI  , etc. have launched " small model " products with only billions of parameters , which are used in various enterprise applications such as voice assistants, search enhancement, and data analysis [6] .

In addition to the small model, other points worth mentioning

IBM 's Watsonx Orchestrate is similar to an enterprise-level AI agent App Store , providing or integrating out-of-the-box AI agent components. Customers can build, assemble or run new AI agents in minutes for rapid deployment.

IBM emphasizes the importance of data quality to generative AI  , a view that is widely shared in the industry, especially in the field of engineering practice. " Small model +  high-quality data +  tool chain integration" is enough to build an enterprise-level AI that is " sufficient, easy to use, and controllable " .