dLLM is an open-source framework that makes diffusion language models simple to train, evaluate, and experiment with. The GitHub project has collected around 2,710 stars and was last actively updated in 2026. Diffusion LLMs generate text by iteratively denoising a whole sequence rather than predicting one token at a time, and dLLM packages this emerging research area into a clean, unified codebase instead of scattered paper repositories.
The framework provides a common pipeline for discrete diffusion language models: data handling, training loops, sampling strategies, and evaluation under one consistent interface. Researchers can reproduce published approaches, swap components such as masking schedules and samplers, and benchmark models fairly against each other. It is built in PyTorch and documented as a companion to the project’s research paper, with the goal of lowering the barrier for anyone who wants to explore non-autoregressive text generation seriously.
Diffusion language models have attracted attention because they can refine an entire draft in parallel and offer different quality-versus-latency trade-offs than autoregressive models, but progress was slowed by fragmented, paper-specific code. dLLM’s contribution is infrastructure: shared data pipelines, training code, and evaluation harnesses that make results comparable across methods. For the research community, that kind of common ground is often what turns a curiosity into a fast-moving subfield.
dLLM is free and open source under the Apache-2.0 license. There is no commercial version; you provide your own training and inference compute.
NLP researchers and ML engineers interested in diffusion-based text generation who want a solid, shared starting point instead of reimplementing each paper from scratch.