nace-ai/drex-decision-models is a repository publishing Drex-family decision models that, instead of continuing a sentence, return probabilities for predefined options, together with how to run them. According to the README, the format takes the input state and a named multiple-choice question and returns a probability for each option. The repository description, model pages, and sample request files are the materials for checking this input/response format and the installation paths. It covers the models and execution paths published in the repository, not a new release of a specific version.
The models introduced include Drex Max, a 27B dense multimodal model, and Drex v2.0, a 35B mixture-of-experts model. The repository also guides through a Python reference implementation, dedicated llama.cpp and Ollama forks, and the Kapitsa and MLX runners. Per-model weights and license information are provided as well; the repository's primary language is Python and the repository license is Apache-2.0. The key point is that the option-probability return method, which differs from text generation models, can be examined across multiple execution paths.
FACTS does not record a latest release, and no performance benchmark figures are presented either. Model scale and execution method alone cannot tell you the actual choice accuracy or stability. The published description is useful for understanding the input format and the execution paths, but it is insufficient as grounds for comparing model performance or deciding whether to apply it in practice.