LingDT-Decision-1.0 takes the opposite road from chat models. On October 22, 2026, Ant Digital Technologies released this decision model built for fast industrial decision-making, in 4B and 27B versions, now live on its DTMaaS platform with a limited-time free trial for developers. It does not write essays or hold conversations; its only job is to return accurate judgments among a limited set of options, quickly. In tests across customer service, risk control and finance, Ant Digital says the model cut token consumption by more than 90 percent versus general large language models on the same tasks, responded about ten times faster, and still judged better.
In enterprise workloads, the expensive part is judgment
The AI judgments an enterprise makes every day far outnumber its generation tasks. When a customer asks why an order has not shipped, the system must judge the issue type, the ticket's urgency, the owning department and whether a human should take over. Risk teams score transactions; finance flows classify and grade. Running these through a general LLM means paying for word-by-word reasoning and generated text — a long, costly chain with less stable outputs. A decision model cuts the generation step entirely: no prose, just judging, single-choice selection, classification and scoring among given options, which fits high-frequency work like intent detection, tool selection and rule-based sorting.
Native multi-select and sequential decisions headline this release
Against comparable decision models, LingDT-Decision-1.0 emphasizes two new abilities. Native multi-select lets one request flag several applicable intents at once instead of splitting a compound question into rounds of single choices. Sequential decision-making organizes judgments as a directed acyclic graph, where earlier conclusions constrain later options, closer to how linked business processes actually run. On evaluation, Ant Digital says the model outperformed Microsoft-Decision-1, Jev 1.13 and similar products across 23 open benchmarks, with the 27B version posting the highest accuracy among models of its parameter scale. Its published business-test details include risk-identification accuracy in smart customer service rising from 76 to 89 percent, and hallucination recall in finance scenarios improving from 44 to 70 percent.
Other decision models launched the same day took a different route — open-sourcing option-scoring models — while Ant Digital ships this one on its own MaaS platform, pitched at dropping into existing customer service, risk and finance flows. Its limits are equally clear: it judges and does not generate, so complex reasoning and content production still belong to general LLMs. The two are a division of labor, not a replacement. The calculation that matters for enterprises is how many high-frequency judgments are currently being forced through a large model; if the volume is big enough, routing that traffic to a dedicated decision model saves more in tokens and latency than any leaderboard position.
