StartLux-Decision debuts RL-trained calibrated decision models, 27B scores 72.8

startlux-models · hf · 2026-10-15

StartLux-Decision is a new family of decision models (0.8B to 27B) on Hugging Face, built for efficient, calibrated deliberative decisions via single-pass prediction and confidence-guided reasoning.

Training: a three-stage pipeline centered on RLCD² reinforcement learning, teaching models when to deliberate and when to answer directly.
Evaluation: benchmarked on the self-introduced Decision Index and DecisionPivot; the official chart reports the 27B model scoring 72.8 on DI 0.3, with 9B at 70.0, 4B at 67.9, and 0.8B at 53.8, claiming top scores per size class among decision models.
Positioning: aimed at classification, routing, and decision tasks rather than general chat.

Original post →

StartLux-Decision debuts RL-trained calibrated decision models, 27B scores 72.8

More from Models