With the popularity of autonomous execution agents like OpenClaw, it marks that AI applications have accelerated from "dialogue interaction" to "task execution." While enterprises are embracing this trend rapidly, they also face multiple challenges such as wasted computing power and security compliance. How to truly achieve large-scale and sustainable deployment of intelligent agents has become a core topic of concern in the industry.

On March 26th, at the Future Industry Innovation Forum of the Zhongguancun Forum, Zhang Peng, General Manager of the Large Model Technology Innovation Department at Ant Digital, stated in his speech that the rise of OpenClaw will bring a revolutionary shift in the enterprise-level AI paradigm, driving the deployment of large models in industrial scenarios from "parameter competition" to "Token efficiency competition."

image.png

Caption: Zhang Peng, General Manager of the Large Model Technology Innovation Department at Ant Digital, delivers a speech at the Zhongguancun Forum

The rapid adoption of OpenClaw-like intelligent agents reflects the market's demand for autonomous execution agents. However, their deployment faces significant challenges in real industrial environments: due to a lack of deep understanding of industry rules and business processes, intelligent agents often repeatedly call tools when executing complex tasks, resulting in token consumption far exceeding actual output. According to reports, in some high-frequency scenarios, the token consumption cost of OpenClaw can be tens or even hundreds of times that of integrated agents, making this high-input-low-output model face sustainability issues in industrial large-scale application.

"The core issue in the second half of the large model industrialization is not the competition of model parameter scale, but the continuous improvement of per-Token efficiency," Zhang Peng said. He believes that enterprises should choose AI solutions that combine large and small models based on actual scenarios and needs, achieving higher business value with lower computing costs.

Take the financial scenario as an example. This field requires processing massive high-frequency low-latency tasks every day—quickly identifying intent, extracting key information, retrieving and sorting, etc. These tasks require high concurrency, fast response, and high accuracy. Traditional industry inference large models are powerful, but in these scenarios, they are like "using a bull to kill a chicken," with high costs, slow response, and resource waste.

"What the industry really needs is an AI solution that ensures professionalism, rigor, and compliance while achieving the optimal cost-effectiveness and response speed," Zhang Peng said. He believes that large-parameter models perform better in complex reasoning and in-depth analysis, while small-parameter models offer lower latency and higher cost-effectiveness in high-frequency small-task scenarios. An integrated approach combining large and small models is needed to more efficiently and cost-effectively solve real-world scenarios.

At the Zhongguancun Forum, Ant Digital released a lightweight financial-specific model called Ling-DT-Fin-Mini-2.5, which is the first model in the Ling DT series of large models. According to the introduction, Ling DT Fin Mini 2.5 is a lightweight MoE model based on the latest hybrid linear attention architecture of Ling 2.5, optimized for high-concurrency, low-latency tasks in the financial sector. It maintains professional depth while compressing the inference cost to a level suitable for large-scale deployment. Compared to mainstream general models of similar capability in the industry, its inference speed is 100% faster, and the hardware cost for processing the same amount of tasks is significantly reduced, bringing practical cost-saving and efficiency-enhancing value to financial institutions.

In fact, as AI agents accelerate their penetration into core industrial scenarios and perform real tasks, the combination of large and small models has become an industry trend. Recently, OpenAI launched two small models, GPT-5.4mini and nano, focusing on low latency and high cost-effectiveness, serving as the main intelligent agents at the execution level.

Zhang Peng stated that technological development will ultimately return to the rational requirements of industrial efficiency. In the next stage of competition, Token efficiency will become the core indicator for measuring the value of enterprise-level AI. Ant Digital will continue to focus on enterprise-level AGI, further launching the enterprise version of the Ling DT large model and its industry versions, accelerating the large-scale deployment of intelligent agents in complex enterprise scenarios.