StartLux Launches Open-Source Decision Model StartLux-Decision in Five Sizes
On Oct 2, 2026, Shanghai-based AI company StartLux (Origin Star) released its dedicated decision model StartLux-Decision and open-sourced the code and model weights. The initial release includes five parameter sizes: 0.8B, 2B, 4B, 9B, and 27B, targeting repeated judgment tasks during agent execution.
The model takes the current state and candidate options as input, supporting multiple-choice, true/false, and rating questions. It can return judgments and option probabilities for multiple queries in a single forward pass, without generating text token by token. For example, one customer service ticket can be simultaneously assessed for its assigned team, urgency, and impact level.
In the independent Decision Index 0.2.1 benchmark, StartLux-Decision-27B scored 63.88, ahead of Jev 1.13's 57.91, and outperformed in 31 of 38 benchmarks. Using the seven-task evaluation suite from Shanghai AI Laboratory's Intern-Decision, the 27B version achieved an average accuracy of 91.82%, higher than Jev's 88.74% and Intern-Decision-4B's 90.02%. For latency, the 4B version averaged 26 milliseconds for three questions on a single H200 with BF16 precision.
StartLux said the model was developed and validated in just three days, powered by its Auto Research approach under construction. Part of the company's RSI (Recursive Self-Improvement) system, Auto Research involves AI in the hypothesis-experiment-validation loop, analyzing failures, constructing data, and running training and evaluation, while human researchers retain control over goals and key trade-offs.
The company has released the model weights, related code, and benchmark results. StartLux plans to combine the decision model with general-purpose models, agents, and runtime modules for complete on-premise intelligent solutions.
