d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation 🚀
This repository contains the d3LLM-Dream model, an ultra-fast diffusion language model introduced in the paper d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation.
- 📄 Paper: arXiv:2601.07568
- 👉 Code repo: https://github.com/hao-ai-lab/d3LLM
- 🌐 Blog: https://hao-ai-lab.github.io/blogs/text-diffusion/
- 🕹️ Demo: https://d3llm-team.github.io/
Model Description
d3LLM-Dream is an ultra-fast diffusion language model that achieves high generation speed while maintaining competitive performance. It strikes a balance between accuracy and parallelism by using pseudo-trajectory distillation during training and entropy-based multi-block decoding during inference.
Key Features
- 🚀 High throughput: 4.5× faster than autoregressive models (Qwen-2.5-7B) on H100 GPU, 2.5× faster on A100 GPU. Achieves 235.34 tokens/s on H100 on GSM8K-CoT.
- 📊 High AUP: Optimized for Accuracy Under Parallelism across benchmarks.
- 🔧 Specialized: Optimized for coding and math reasoning tasks.
Usage
For more chat examples and evaluation scripts, visit the official repository.
Citation
@inproceedings{ICML'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},
booktitle = {Proceedings of the 43rd International Conference on Machine Learning (ICML)},
pages = {to appear},
year = {2026}
}