Ant Digital Technologies releases decision model LingDT-Decision-1.0, 27B version ranks first in overall score

Zhidongxi October 22 report: Today, Ant Digital Technologies released the decision model LingDT-Decision-1.0. The model offers two versions, 4B and 27B, purpose-built for high-frequency decision tasks in industrial AI scenarios such as customer service, risk control, and finance, and debuts native multi-select and ordered decision (DAG) capabilities.

Currently, the model is live on the Ant Digital Technologies DTMaaS platform, and developers can try it free for a limited time.

Decision models do not generate long-form text; they directly return options, scores, or judgment results, and can be used for intent recognition, ticket routing, tool selection, model routing, and risk judgment, among others. This type of model has recently drawn sustained attention after Jev went viral. Following Jev, vendors such as Microsoft and Perplexity have successively launched their own decision models.

LingDT-Decision-1.0 and Jev both target structured decisions, but with different emphases. Jev is more general-purpose, mainly aimed at the general decision layer in developer and software systems, and is commonly used to route requests, select tools, arrange Agent workflows, or trigger manual review; LingDT-Decision-1.0 is more specialized, mainly aimed at composite decisions and multi-step workflows in enterprise businesses such as customer service, risk control, and finance.

In terms of model capabilities, Jev mainly completes selection, scoring, or true/false judgment through preset options. Among these, Choice typically returns the single highest-probability option, and multiple judgments can also be achieved by posing multiple questions in parallel. LingDT-Decision-1.0, by contrast, can identify multiple requests simultaneously and adjust the processing order according to the dependencies among tasks.

Tests by the Ant Digital Technologies team show that when handling the same tasks in scenarios such as data annotation, LingDT-Decision-1.0 compared with DeepSeek-V4-flash can reduce Token consumption by 97%, improve response latency by about 10x, while maintaining better performance.

Across 23 open-source evaluation benchmarks, LingDT-Decision-1.0's overall score surpassed popular decision models such as Microsoft-Decision-1 and Jev 1.13.0, and the 27B version achieved the highest accuracy among models of the same parameter count.

1. Token consumption reduced by over 90%, covering high-frequency tasks such as customer service and risk control

Enterprise businesses contain a large number of tasks that do not require generating a complete answer but do require quickly completing a judgment.

Traditional large language models typically need to first generate a structured answer, which is then parsed and validated by downstream programs. LingDT-Decision-1.0 can instead complete judgments directly within a limited set of options, reducing the impact of unstable output formats.

Test data published by Ant Digital Technologies shows that in data annotation scenarios, LingDT-Decision-1.0 compared with DeepSeek-V4-flash reduces Token consumption by about 97% and improves processing latency by about 10x; in intelligent customer service scenarios, compared with Qwen-3.7-flash, it reduces Token consumption by 90%, improves processing latency by about 1x, and raises risk identification accuracy from 76% to 89%.

In financial application scenarios, LingDT-Decision-1.0 can be used for intent recognition, tool invocation, and hallucination detection. Tests show that processing latency for related tasks improves by about 5x, the exact match rate of tool combinations improves by about 2.6x, and hallucination recall rises from 44% to 70%.

Test results show that LingDT-Decision-1.0 exhibits low Token consumption, low latency, and relatively high accuracy across scenarios such as customer service, risk control, and finance. The same model can also cover multiple types of decision tasks, making it easy for enterprises to reuse across different business workflows.

2. Supports native multi-select and ordered decision, 27B version ranks first in overall score

Ant Digital Technologies states that in addition to supporting basic capabilities such as judgment, single-select, classification, and rating, LingDT-Decision-1.0 also debuts native multi-select and ordered decision capabilities, adapting to complex, multi-step decision needs in enterprise businesses.

Native multi-select capability allows the model to hit multiple candidate options simultaneously within one task, suitable for multi-label recognition, composite intents, and tool combinations. For example, when a user says "I want a refund, to cancel renewal, and to get an invoice issued, but do not cancel the account," the model can simultaneously identify multiple requests such as refund, cancel renewal, and invoice issuance.

Ordered decision can complete multi-step judgments according to the dependencies among tasks. For example, a refund needs verification first, canceling renewal can be executed immediately, and invoice processing may need to wait for the refund result. Compared with single-item decisions that return only one option, this approach is closer to real processing workflows in enterprise businesses.

Across 23 mainstream open-source evaluation benchmarks, the 27B version of LingDT-Decision-1.0 ranks first in overall score. The score consists of 90% single-select results and 10% multi-select results. Ant Digital Technologies states that LingDT-Decision-1.0's accuracy exceeds decision models such as Microsoft-Decision-1, Jev 1.13.0, and Fastino/GLiDE.

Among these, on multi-select tasks, the 4B and 27B versions' accuracy exceeds 16 decision models including Microsoft-Decision-1, Jev 1.13.0, and Star-lux; on single-select tasks, the two versions are also in the top tier.

Decision models provide a new solution for high-frequency tasks such as intent recognition, task classification, tool selection, and risk judgment. Because they do not need to generate long-form text, they can directly return structured results, thereby reducing Token costs and improving response speed.

With this release of LingDT-Decision-1.0, Ant Digital Technologies further places decision models into enterprise business workflows such as customer service, risk control, and finance for validation. Native multi-select and ordered decision capabilities also enable the model to handle composite intents and multi-step tasks. As more decision models enter real business operations, their performance and applicable boundaries will be further tested.