LLaDA is the official open-source PyTorch implementation of Large Language Diffusion Models, a research project exploring whether diffusion — the technique behind modern image generators — can power large language models as an alternative to the usual autoregressive approach. Published on GitHub, the repository has attracted about 4,003 stars and was updated in 2026 as the work continues to evolve.
Instead of generating text strictly one token at a time from left to right, LLaDA learns to reconstruct clean text from progressively masked versions of it. At generation time the model starts from a fully masked sequence and iteratively unmasks tokens, refining the whole answer in parallel. The repository provides the training and inference code behind the paper, together with the released LLaDA model checkpoints, so researchers can reproduce results, probe the strengths of diffusion-based generation — such as bidirectional context and natural infilling — and compare it fairly with autoregressive baselines of similar scale. The release attracted wide attention because it showed diffusion models can reach quality competitive with similarly sized autoregressive systems on standard benchmarks and instruction-following tests.
LLaDA is free and open source: the code and research artifacts are published on GitHub for the research community at no cost. Running the models requires your own GPU hardware or cloud compute.
LLaDA is for machine learning researchers and advanced practitioners interested in alternatives to autoregressive language modeling — anyone studying diffusion models for text, evaluating new generation paradigms, or building on an influential open research release.