startluxlabs/startlux-decision is a collection of public decision models that outputs a calibrated probability for each option in a single forward pass. Model sizes range from 0.8B to 27B, and the latest release is v1.1.0. The repository description and README state that it presents probabilities instead of generating long answers, and that AutoThink models add short reasoning when confidence is low.
Questions are entered with a type such as choice, yes·no, or scale. Instant models use bidirectional attention in the prompt to output the answer in one shot. AutoThink models answer in a single pass when the top probability exceeds the default 0.6, and otherwise generate short reasoning and then re-read the answer. It supports vLLM, llama.cpp, and MLX as inference backends, and ships 8-bit and 4-bit weights for GGUF and MLX. The project is written in Python and follows the Apache-2.0 license.
The README reports the following Decision Index 0.3 scores for v1.1: 27B AutoThink 73.37, 27B Instant 72.83, 9B AutoThink 70.45, and 9B Instant 69.96. The DecisionPivot scores in the same order are 86.9, 75.4, 59.9, and 47.5. However, the README does not show how these scores were computed or how to reproduce them. It is therefore more appropriate to read the numbers as limited to the published evaluation results, rather than as evidence that settles performance between models.