Ant Digital Unveils Bai Ling Enterprise Financial LLM: Speed and Efficiency Redefined

The rise of autonomous execution agents like OpenAI's Agent marks a significant acceleration for AI applications, shifting from simple "dialogue interaction" to complex "task execution." While businesses are quickly embracing this trend, they also face challenges like computational waste and security compliance. Achieving large-scale, sustainable deployment of intelligent agents has become a central industry focus.

On March 26th, at the Zhongguancun Forum's Future Industry Innovation Forum, Zhang Peng, General Manager of the Large Model Technology Innovation Department at Ant Digital, stated in his speech that the emergence of such agents will revolutionize the enterprise AI paradigm. This shift will move industrial deployment of large models from a "parameter scale competition" to a "Token efficiency competition."

Zhang Peng, General Manager of Ant Digital's Large Model Technology Innovation Department, speaking at the Zhongguancun Forum.

The rapid adoption of intelligent agents reflects strong market demand for autonomous task execution. However, deploying them in real industrial settings presents major hurdles. Often lacking deep insight into industry rules and workflows, these agents can make repeated, inefficient tool calls when handling complex tasks. This leads to Token consumption that far exceeds the actual value created. Reports indicate that in some high-frequency scenarios, the cost for agent execution can be tens or even hundreds of times higher than integrated solutions, raising sustainability concerns for widespread industrial use.

"The core challenge in the next phase of large model industrialization is not competing on parameter count, but continuously improving efficiency per Token," Zhang Peng noted. He believes companies should choose AI solutions that strategically combine large and small models based on specific scenarios and needs, achieving greater business value with lower computational cost.

Consider the financial sector as an example. This field processes massive volumes of high-frequency, low-latency tasks daily—quickly identifying user intent, extracting key data, retrieving and sorting information. These tasks demand high concurrency, fast response times, and pinpoint accuracy. While traditional, heavy-duty inference models are powerful, using them for such scenarios is often overkill—like "using a sledgehammer to crack a nut"—resulting in high costs, slow speeds, and wasted resources.

"What the industry truly needs is an AI solution that guarantees professionalism, rigor, and compliance while delivering optimal cost-effectiveness and response speed," Zhang Peng explained. He emphasized that large-parameter models excel at complex reasoning and in-depth analysis, while smaller models offer lower latency and better cost efficiency for high-frequency, simpler tasks. A hybrid approach that leverages both is key to solving real-world problems more efficiently and economically.

At the Zhongguancun Forum, Ant Digital launched a lightweight, finance-specific model named Ling-DT-Fin-Mini-2.5, the first in its Ling DT series. This model is a lightweight Mixture-of-Experts (MoE) architecture based on the latest Ling 2.5 hybrid linear attention framework, optimized for high-concurrency, low-latency financial tasks. It maintains professional capability while reducing inference costs to a level suitable for mass deployment. Compared to mainstream general-purpose models of similar capability, it offers a 100% faster inference speed and significantly lower hardware costs for processing equivalent task volumes, delivering tangible cost savings and efficiency gains for financial institutions.

Indeed, as AI agents move deeper into core industrial processes to perform real work, the strategy of combining large and small models is becoming an industry standard. Recently, OpenAI itself released two smaller models, GPT-4o mini and nano, focusing on low latency and high cost-effectiveness to serve as primary execution-level agents.

Zhang Peng concluded that technological progress must ultimately align with the rational demands of industrial efficiency. In the next stage of competition, Token efficiency will become the core metric for evaluating enterprise AI value. Ant Digital will remain focused on enterprise-grade AGI, planning to release an enterprise edition of the Ling DT large model and industry-specific versions, accelerating the large-scale deployment of intelligent agents in complex business environments.