JetBrains, the company behind the IntelliJ IDE, has said that developer workflows could be disrupted next year by Diffusion Large Language Models (d-LLMs) replacing the dominant autoregressive (AR) models. These models will offer several benefits including out-of-order generation, bi-directional context, flexibility in editing, coordinated multi-region updates, and potential speed improvements.
The core difference between d-LLMs and AR-LLMs is that the prior are capable of non-sequential generation. Developers work non-sequentially, they focus on editing and refactoring their code in an iterative manner rather than typing complete functions in one sequence. AR-LLMs generate code token-by-token in a strict left-to-right sequence whereas d-LLMs look at past and future context and make edits directly and plan token generation more globally. This aligns better with the non-linear nature of writing code we see from human developers.
While d-LLMs hold promise, there are significant drawbacks standing in their way right now. Right now, the best quality output from these models is when they unmask only one token per step, but this slows them down to AR model speeds. Furthermore, when you push a d-LLM, it may output incoherent text like repetition, early termination, and malformed syntax. Right now, state-of-the-art d-LLMs show mixed performance against strong AR baselines.
Despite the issues with d-LLMs, there are some currently useful niches for them. These include code completion with context editing where missing parts are filled, not just extending text; refactoring less rigid code blocks; and structured text tasks and reversal problems that rely heavily on bi-directional context.
According to JetBrains, we could see d-LLMs become more prominent as researchers resolve the quality-efficiency trade-offs. The potential faster code generation could make them the core of future coding assistants that feel like “principled, code structure-aware, programming collaborators.”
We could see this new generation of LLMs show up in AI coding editors such as Visual Studio Code, Cursor, and Windsurf sometime next year. The rise of these AI editors has led to a phenomenon called Vibe Coding where AI handles most of the code generation. The introduction of d-LLMs could further help people rely on AI to write correct code.