JetBrains AI

Diffusion models, and in particular diffusion large language models (d-LLMs), operate differently from current coding assistants. Unlike autoregressive models, which generate token by token and line by line in a strict left-to-right sequence, d-LLMs condition on both past and future context.