This repository contains the full merged 8B model weights of WINO+, initialized from MMaDA-8B-MixCoT.

WINO+ is a trajectory-injection method for diffusion large language models. It transfers the verified denoising order discovered by WINO into model parameters, enabling faster and higher-quality generation.

The model corresponds to the WINO+ method described in Roll Out and Roll Back: Diffusion LLMs are Their Own Efficiency Teachers.