The repository presents the implementation of Multiplex Thinking, a reasoning mechanism that utilizes token-wise branch-and-merge techniques. This approach allows for efficient multi-pattern reasoning without sacrificing the compactness of token representation. The codebase relies on several established open-source projects, acknowledging the contributions of their original authors. To get started, users are encouraged to use Docker for a consistent setup, although there’s also an option for Conda with a provided environment specification file. Key package versions required for operation include sglang version 0.4.9.post6 and transformers version 4.54.0. The setup process involves running a script that installs dependencies and ensures the correct library versions are in use. For training and evaluation, the repository includes specific scripts. Users can train models with parameters such as batch size (128), maximum token length (32,768), and multiplex width (3). The evaluation process is also streamlined with a dedicated script. Model weights can be accessed on Hugging Face, making it easier for researchers to leverage this work in their own studies. Proper citation of the associated paper is recommended for those who benefit from the findings.