Ant Digital Technologies Launches LingDT-Decision-1.0 Decision Model

The model comes in 4B and 27B versions and is available on Ant Digital Technologies' DTMaaS platform, where developers can try it free for a limited time, according to QbitAI.

Ant positions LingDT-Decision-1.0 for high-frequency enterprise decisions such as intent classification, tool selection, rule sorting and result scoring. Unlike large language models that generate text token by token before producing a structured answer, the decision model chooses directly among defined options, which Ant says shortens the process, lowers token costs and reduces format problems from output parsing and validation.

In tests across customer service, risk control and financial applications, Ant said LingDT-Decision-1.0 cut token consumption by more than 90% and improved response time tenfold compared with large language models on the same tasks, while delivering better performance.

The company provided specific comparisons from its business scenarios. In data annotation, token consumption fell about 97% and processing speed rose about tenfold compared with DS-V4-flash. In customer service, token use dropped 90% and processing speed roughly doubled compared with Qwen-3.7-flash, while risk identification accuracy rose from 76% to 89%. In financial applications covering intent recognition, tool invocation and hallucination detection, processing speed increased about fivefold, exact tool-combination matching improved about 2.6 times, and hallucination recall rose from 44% to 70%.

Ant also said LingDT-Decision-1.0 is the first decision model to natively support multi-select and ordered decision (DAG) capabilities, beyond basic judgment, single-select, classification and rating. In a request in which a user asks to refund an order, cancel a subscription, issue an invoice but not close an account, single-select may capture only the main refund request, while multi-select can identify all demands and ordered decision can determine dependencies, such as verifying the refund first and waiting for its result before handling the invoice. Ant said these functions better fit multi-step workflows in finance, risk control and customer service.

On the 23 open-source evaluation benchmarks, Ant said the 27B version ranked first in composite score, calculated as 90% single-select score plus 10% multi-select score, exceeding Microsoft-Decision-1, Jev 1.13.0 and Fastino/GLiDE. On multi-select tasks, the 4B and 27B versions outperformed Microsoft-Decision-1, Jev 1.13.0 and Star-lux in a comparison that included 16 decision models. On single-select tasks, the model was in the first tier, and Ant said the 27B version achieved the highest accuracy among models of the same parameter size.

QbitAI said the article was provided by Ant and republished with authorization, with the views belonging to the original author.