Chinese Decision models produce decision probabilities in a single pass

Researchers report releasing three Chinese Decision models that produce probabilities for predefined choices in a single forward pass, without generating an answer token by token. Built on Qwen3.5 0.8B, 2B and 4B models, they support Boolean judgments, categorical choices and ordinal scores. The preprint reports 85.97% overall accuracy for the 4B model on the 307,900-decision CJ-Bench, with 1.00% expected calibration error. Each model includes a temperature setting fitted on validation data and Q8_0 quantization. All comparisons rely on single training runs and imported baselines; the global temperature setting improves pooled calibration while worsening some domains.