Ant Digital Technologies, a division of Ant Group specializing in enterprise technologies, has introduced the LingDT-Decision-1.0 model. This model is specifically designed for structured business decision-making, rather than free text generation. The model is available in two versions—4B and 27B parameters—and is hosted on the company's DTMaaS platform, where developers can test it for free for a limited time.
The core idea was to isolate a specific type of task from the general field of large language models. Many corporate operations boil down to selecting from predefined options. As an example, Ant Digital Technologies provided a customer service scenario: a message stating that an order has not been shipped within three days requires the system to determine whether it is a refund request, a reminder, or a complaint, what its urgency level is, which department should handle it, and whether human intervention is required.
Standard large language models typically operate by reasoning one token at a time, generating text that is then parsed to derive a decision. The company notes that this process is lengthy, costly, and more prone to errors. The decision-making model is designed to directly evaluate options within a given set of choices. This makes it suitable for high-frequency steps in agent workflows, such as intent recognition, tool selection, rule classification, and evaluation.
According to tests conducted by Ant Digital Technologies in customer service, risk control, and financial scenarios, LingDT-Decision-1.0 reduced token consumption by over 90% compared to large language models when performing similar tasks. Furthermore, it responded up to 10 times faster and demonstrated superior results. Beyond simple 'yes' or 'no' judgments, single-option selection, classification, and rating, the company claims this model is the first to support native multiple-choice and sequential decisions organized as a Directed Acyclic Graph (DAG), aligning with multi-stage business processes.
Across 23 open-source benchmarks, the company stated that LingDT-Decision-1.0 outperformed other decision-making models, including Microsoft-Decision-1 and Jev 1.13, with the 27B version achieving the highest accuracy among models of its size. It should be noted that these results are proprietary to the company and have not undergone independent verification.
The release of this model does not mean the discontinuation of general large language models. Instead, the model is intended to work in parallel with them, taking on many small, repetitive judgments within a corporate agent so that larger models can focus on free-form generation and reasoning. For teams building agents at scale, this separation can significantly impact cost and latency.