A language model that generates multiple tokens in parallel using diffusion, then refines them iteratively.