Chinese AI Startup's Decision Model Tops Hugging Face Global Ranking

On October 17, Hugging Face updated its Decision Index 0.3.2 leaderboard, with StartLux-Decision-v1.1-27B Instant, developed by Shanghai-based AI startup StartLux, taking the top spot with a score of 69.1. The result was nearly 3 points higher than the runner-up and made it the only model to surpass 69 points among 173 publicly reproduced models evaluated.

The model is a rapid iteration following StartLux's first-generation decision model released in late September. The company also introduced two technical routes — Instant for high-speed judgment and AutoThink for on-demand reasoning — along with the RLCD² reinforcement learning training method and the DecisionPivot evaluation benchmark. All model weights and inference code have been open-sourced.

Beyond the Instant model, StartLux-Decision-27B-AutoThink scored 73.37 in professional tests. The company also reported that its 4B-parameter AutoThink model achieved 68.21, outperforming several larger mainstream hosted models. As of October 17, cumulative downloads of the series had reached nearly 120,000.

Decision models are positioned as the core judgment center for AI agents, handling task scheduling, tool invocation, and information verification. StartLux said the Instant model is designed for high-frequency, lightweight classification and routing tasks, while AutoThink activates deeper reasoning in complex and uncertain scenarios. Its RLCD² method uses reward-and-punishment mechanisms to curb ineffective reasoning; according to the company, the 9B model's rate of autonomously terminating invalid reasoning rose from 45.9% to 93.9%, with average reasoning length reduced by 65%.