d3LLM/d3LLM_Dream_Coder

d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation 🚀

d3LLM-Dream-Coder is an ultra-fast diffusion language model introduced in the paper d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation. It is built on Dream-org/Dream-Coder-v0-Instruct-7B.

Model Description

d3LLM (pseuDo-Distilled Diffusion Large Language Model) is a framework designed to strike a balance between accuracy and parallelism in diffusion LLMs. It achieves up to 10× speedup over vanilla diffusion models like LLaDA/Dream and 5× speedup over autoregressive (AR) models.

The model utilizes two primary innovations:

• Pseudo-Trajectory Distillation: A training method that teaches the model which tokens can be decoded confidently at early steps.
• Entropy-Based Multi-Block Decoding: An inference strategy using a KV-cache refresh mechanism to maintain accuracy while maximizing parallelism.

Resources

• Paper: d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation
• Repository: https://github.com/hao-ai-lab/d3LLM
• Blog: https://hao-ai-lab.github.io/blogs/text-diffusion/
• Demo: https://d3llm-team.github.io/

Usage

For detailed usage instructions, evaluation scripts, and training code, please refer to the official GitHub repository. Since the model uses a custom architecture, ensure you have transformers==4.49.0 installed and use trust_remote_code=True when loading the model.

Citation

@article{arxiv'26:d3llm,
  title   = {d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation},
  author  = {Yu-Yang Qian and Junda Su and Lanxiang Hu and Peiyuan Zhang and Zhijie Deng and Peng Zhao and Hao Zhang},
  journal = {ArXiv preprint},
  volume  = {arXiv:2601.07568},
  year    = {2026}
}