Masked Diffusion Language Models (MDLM) are a novel type of language model that utilizes a masking and diffusion mechanism to generate high-quality text data. MDLM improves upon existing diffusion models through advanced training methods and a simplified objective function, achieving new state-of-the-art performance in language modeling benchmarks and approaching the perplexity of autoregressive models. Key advantages of MDLM include an efficient sampling method, support for generating text of arbitrary length, and strengths in long-range dependencies and controlled generation.