Ant Digital Technologies has released 4B and 27B versions of LingDT-Decision-1.0 for business tasks requiring choices among defined options, with its 27B version topping an aggregate evaluation of 23 open-source benchmarks.
The 27B model posted higher accuracy than Microsoft-Decision-1, Jev 1.13.0 and Fastino/GLiDE, and led other models of similar size.
LingDT-Decision-1.0 selects from preset options rather than producing open-ended prose. It supports outputs with multiple choices and can apply decisions in dependency order.
Across customer service, risk management and finance tests, token use was more than 90% lower, while response times improved by up to a factor of 10. Customer-service risk-identification accuracy moved from 76% to 89%; finance hallucination recall went from 44% to 70%.
Perplexity offers another example of the category: its two decision models charge $0.01 or $0.02 per million input tokens, with output tokens free, according to earlier EconoTimes coverage.
