A survey of how modern diffusion language models are actually built, and why parallel token generation is starting to look like a real alternative to autoregression

Mainstream LLMs generate autoregressively, one token at a time from left to right. That leaves parallel GPU hardware underutilised during inference - a bottleneck that has grown acute as progress shifts toward post-training and inference-time compute - and it commits irreversibly to each token with no way to revise an earlier mistake. Discrete diffusion models address both by generating tokens in parallel, starting from a noisy or masked sequence and iteratively refining the whole output over a series of denoising steps. The parallel process enables speeds of thousands of tokens per second, provides a new axis for inference-time compute scaling, and supports built-in error correction and controllable generation. Over the past two years discrete diffusion has scaled to billions of parameters across biological sequence modelling and natural language, with open-weight and commercial releases from leading labs. This article is a survey of the key ideas in discrete diffusion and a high-level introduction to how modern diffusion LLMs are built, covering foundations, core algorithmic and architectural building blocks, applications in biology and language, and implications for future LLM scaling laws. The honest limit: this is a survey, not a new result, so it reports the state of a field rather than advancing it. The thousands-of-tokens-per-second claim describes the generative process in principle, and real serving is bounded by the denoising steps required to reach a given quality. https://arxiv.org/abs/2610.33391