AI News Feed
Market watch
Woot Cuts Apple Magic Keyboards for iPad Air and iPad Pro to as Low as $129Two PHP flaws put wide swaths of web at risk, forcing site shutdownsPrediction Markets Upstage Election Polls as Kalshi, Polymarket Draw Betting UseSpotify Touts Agent Strategy as Q3 Users Reach 788 Million; Confirms Family Plan ChecksSwitch 2 Apparently Jailbroken With SmileBASIC 4 and OpenAI CodexLe Monde Reports 96% of Coral Reefs 'Critically Endangered', at Risk of Collapse Within 50 Years
Woot Cuts Apple Magic Keyboards for iPad Air and iPad Pro to as Low as $129Two PHP flaws put wide swaths of web at risk, forcing site shutdownsPrediction Markets Upstage Election Polls as Kalshi, Polymarket Draw Betting UseSpotify Touts Agent Strategy as Q3 Users Reach 788 Million; Confirms Family Plan ChecksSwitch 2 Apparently Jailbroken With SmileBASIC 4 and OpenAI CodexLe Monde Reports 96% of Coral Reefs 'Critically Endangered', at Risk of Collapse Within 50 Years
Large Language Models

Ant Digital Releases LingDT-Decision-1.0 Decision Model, Claims Top Scores on 23 Benchmarks

Ant Digital launches LingDT-Decision-1.0, claiming 90% less token use and 10x faster response, Leiphone reports.

AI News Feed · 2026-10-23 00:03

The model comes in 4B and 27B versions. It is available on Ant Digital's DTMaaS platform for a limited free trial, Leiphone reported.

LingDT-Decision-1.0 is designed to make choices from limited options rather than generate long text, the report said. Traditional LLMs in such scenarios usually reason token by token and then generate output text, a longer chain that can be costly and error-prone, according to the report. Ant Digital says decision models are aimed at making fast, accurate judgments rather than producing text, which suits high-frequency business tasks such as intent judgment, tool selection, rule classification and result scoring.

In tests across customer service, risk control and financial scenarios, Ant Digital reported that the model used more than 90% fewer tokens than LLMs on the same tasks and responded 10 times faster.

In data annotation, token use fell about 97% and processing time improved about 10 times compared with DS-V4-flash, according to the company. In intelligent customer service, token use fell 90% and processing time improved about one time compared with Qwen-3.7-flash, while risk recognition accuracy rose to 89% from 76%.

In financial applications covering intent recognition, tool calling and hallucination detection, processing time improved about five times, exact matching of tool combinations rose about 2.6 times, and hallucination recall increased to 70% from 44%, Ant Digital said.

The model adds native multi-select and ordered decision, or DAG, capabilities. Ant Digital says these features fit business processes where a request can contain several tasks that must be handled in a sequence. For example, a customer may ask for a refund, cancellation of renewal and an invoice while asking not to close the account; the model can identify all requests and decide which steps depend on others.

On 23 open-source benchmarks, the 27B version ranked first in a composite score weighting single-select at 90% and multi-select at 10%, exceeding Microsoft-Decision-1, Jev 1.13.0 and Fastino/GLiDE, Leiphone reported. The 4B and 27B versions also beat 16 decision models including Microsoft-Decision-1, Jev 1.13.0 and Star-lux on multi-select tasks, while placing in the first tier on single-select tasks.