Ant Group Digital Technologies releases decision model LingDT-Decision-1.0, offering two versions: 4B and 27B

On October 22, Ant Digital Technologies officially released its decision model LingDT-Decision-1.0. The model is purpose-built for fast decision-making and, when handling structured decision tasks, offers advantages including higher accuracy, faster responses, and lower token costs, making it especially suited to industrial AI scenarios. LingDT-Decision-1.0 is available in two versions, 4B and 27B. The model is now live on the Ant Digital Technologies DTMaaS platform, where developers can try it free for a limited time.

Tests by the Ant Digital Technologies team in real business scenarios such as customer service, risk control, and finance show that, when handling the same tasks, LingDT-Decision-1.0 can reduce token consumption by more than 90% compared with large language models, improve response time by 10x, and deliver better performance.

Compared with other decision models, LingDT-Decision-1.0 is the first to natively support multi-select and ordered decision-making, making it better suited to real enterprise application scenarios. Across 23 open-source evaluation benchmarks, LingDT-Decision-1.0's overall score surpassed popular decision models such as Microsoft-Decision-1 and Jev 1.13. The 27B version of LingDT-Decision-1.0 achieved the highest accuracy among models of the same parameter count.

Designed specifically for decision tasks, token consumption cut by more than 90%, faster and more accurate

In real enterprise applications, there are a large number of high-frequency decision-type tasks.

For example, in a customer service scenario, after receiving the inquiry "This order still hasn't shipped after three days, please handle it as soon as possible," the AI needs to quickly make a series of judgments and decisions: the problem type (refund, expedite, or complaint?), the ticket urgency (does it need priority handling?), the department responsible (which department should the ticket go to?), and the risk judgment (does it need human takeover?).

Traditional large language models have obvious shortcomings in such scenarios. Large language models usually need to reason token by token and then generate output text, resulting in a long pipeline, high costs, and greater susceptibility to errors, with performance that is not stable enough.

Decision models are designed precisely for these pain points. Their goal is not to generate text but to make fast, accurate judgments. As such, they are naturally suited to handling high-frequency decision-type tasks in enterprise operations, such as intent judgment, tool selection, rule classification, and result scoring.

Currently, the LingDT-Decision-1.0 model has been tested in real business scenarios at Ant Digital Technologies, including intelligent customer service, security risk control, and financial applications. Its core approach is to shift large models from "first generating a structured answer, then performing subsequent parsing" to "making judgments directly among limited options." This not only reduces format issues caused by unstable output but also makes the originally cumbersome parsing and validation process simpler.

For example, in a data annotation scenario, compared with DS-V4-flash, token consumption is reduced by about 97% and processing time is improved by about 10x; in an intelligent customer service scenario, compared with Qwen-3.7-flash, token consumption is reduced by 90% and processing time is improved by about 1x, while risk identification accuracy rises from 76% to 89%.

In financial application scenarios, across tasks such as intent recognition, tool invocation, and hallucination detection, processing time is improved by about 5x, the exact-match rate for tool combinations increases by about 2.6x, and the hallucination recall rate rises from 44% to 70%, making risk detection more sensitive.

Overall, LingDT-Decision-1.0 consistently demonstrates the advantages of lower cost, faster response, and more accurate judgment across multiple real-world scenarios. At the same time, it can adapt to multiple types of judgment tasks with a single model, making integration simpler and maintenance less effortful, and therefore easier to reuse and promote across different business scenarios.

Building the next-generation decision model, with two features making their global debut

Compared with other decision models, Ant Digital Technologies' LingDT-Decision-1.0 places greater emphasis on being built for real enterprise application scenarios. In addition to basic capabilities such as judgment, single-select, classification, and rating, LingDT-Decision-1.0 also debuts native multi-select and ordered decision-making (DAG) features, which better adapt to the complex, multi-step decision-making needs of enterprise operations.

In enterprise scenarios, what is more common is not "choose one" but "identify a set of items to handle and execute them in order." For example, when a user says: "I want a refund, to cancel auto-renewal, and to get an invoice issued, but do not cancel my account." Single-select often can only identify the main request of "refund" and easily misses other needs. Multi-select can identify all requests at once; ordered decision-making can further determine: the refund needs verification first, canceling auto-renewal can be executed first, and invoice processing may have to wait for the refund result. Therefore, multi-select and ordered decision-making are closer to real business processes and can better improve accuracy and final handling outcomes.

In terms of native multi-select capability, LingDT-Decision-1.0 allows a single task to match multiple candidates simultaneously, adapting to needs such as multi-label recognition, composite intents, and tool combinations; in terms of ordered decision-making capability, the model can complete multi-step decisions according to dependencies between tasks, getting closer to real business processes in finance, risk control, customer service, and other areas.

Test results covering 23 mainstream open-source evaluation benchmarks show that the 27B version of LingDT-Decision-1.0 ranks first in overall score (90% x single-select score + 10% x multi-select score), with accuracy surpassing popular decision models such as Microsoft-Decision-1, Jev 1.13.0, and Fastino/GLiDE. Among them, on multi-select tasks, the accuracy of the 4B and 27B versions of LingDT-Decision-1.0 exceeds that of 16 decision models including Microsoft-Decision-1, Jev 1.13.0, and Star-lux, demonstrating the advantages of native multi-select capability in multi-label and combined judgment tasks. On single-select tasks, it ranks in the first tier.

(Figure caption: A two-dimensional chart of single-select/multi-select capabilities shows that LingDT Decision 4B/27B belongs to Tier-1 in single-select & judgment capability, surpassing the vast majority of models of the same size; while its "native multi-select capability" leads by a wide margin)