vladimirrott/mojo-dllm is a Mojo implementation that runs the diffusion language models LLaDA-8B and Dream-7B locally. The latest release recorded in FACTS is v0.1.0. The provided material does not separately list anything newly added or changed in this version. According to the repository description and the README summary, the implementation reads quantized GGUF files and performs the iterative generation process of the models.
GGUF is a file format that holds model weights, and this repository builds the inference path in Mojo, from reading GGUF through the tokenizer, matrix operations, GPU kernels, the transformer, and iterative generation. LLaDA-8B and Dream-7B can run on CPU or an NVIDIA GPU, and Python is not needed at runtime. Tests for the model and tokenizer modules are also included in the tree. The implementation is licensed under Apache-2.0, and its core is that it handles the local inference path for diffusion language models in Mojo.
Performance data has to be read with its conditions attached. The comparisons in the README mix different hardware, models, and measurement procedures, so the numbers are hard to compare directly, and the benchmark results have not been independently reproduced. The fact that the repository provides a run path and tests can therefore be confirmed, but actual performance is hard to judge from the presented comparison numbers alone.