Multi-Block Diffusion Language Models (MBD-LMs)

This repository contains the model weights for Multi-Block Diffusion Language Models (MBD-LMs), as presented in the paper Multi-Block Diffusion Language Models.

• Project Page: https://sjtu-deng-lab.github.io/mbd-lms
• GitHub Repository: https://github.com/SJTU-DENG-Lab/mbd-lms
• Inference Engine: https://github.com/SJTU-DENG-Lab/Diffulex

Introduction

Block Diffusion Language Models (BD-LMs) improve diffusion-based text generation with KV caching and flexible-length generation. MBD-LMs extend them from Single-Block Diffusion (SingleBD) to Multi-Block Diffusion (MultiBD), where a running-set of consecutive blocks is decoded concurrently for inter-block parallelism.

By post-training BD-LMs with Multi-block Teacher Forcing (MultiTF) and utilizing an optimized decoding algorithm based on the Block Buffer mechanism, MBD-LMs translate increased decoding parallelism into significant wall-clock acceleration without sacrificing quality.

Citation

If you find this work useful, please cite the paper:

@article{jin2026multi,
  title={Multi-Block Diffusion Language Models},
  author={Jin, Yijie and Xu, Jiajun and Liu, Yuxuan and Xu, Chenkai and Tu, Yi and Li, Jiajun and Tu, Dandan and Yan, Xiaohui and Yu, Kai and Liu, Pengfei and Deng, Zhijie},
  journal={arXiv preprint arXiv:2606.29215},
  year={2026}
}