Diffusion LLM

Most language models write text the way a human types: one word after another, always from left to right. A Diffusion LLM — short for Diffusion Large Language Model, i.e. a large language model based on diffusion — proceeds differently. It starts with a complete but initially meaningless placeholder text and refines it in many small steps until a readable answer emerges at the end. Each step corrects the entire result so far all at once, not just the next word. This is a fundamentally different architecture that brings certain advantages and disadvantages with it.

Why Diffusion LLMs are reigniting the architecture debate

The classic principle of generating text from left to right has a structural disadvantage: the model commits early. If it chooses an unsuitable word at the beginning, it can no longer easily correct this later — it only ever works forwards. A Diffusion LLM does not have this problem. It can revise the entire draft at any time, because it is never permanently fixed from left to right.

This is particularly relevant for tasks where the correct phrasing depends heavily on the end of the sentence. Take translation, for example: sometimes you only know which verb belongs at the beginning of the English version once you’ve reached the end of a long German sentence. A model that thinks about the whole text simultaneously has a fundamental advantage here over one that works forward word by word.

From noise to readable answer — the process

The term “diffusion” originally comes from physics and describes how something spreads and distributes itself — such as ink in water. In AI research, the concept was initially used for images: you take a real image, gradually add noise until nothing is recognizable anymore, and train a model to reverse this process. A Diffusion LLM applies the same principle to text. Instead of pixels, words — or more precisely their numerical representations, so-called tokens — are noised and then restored.

During training, the model learns to reconstruct the original text from a corrupted, randomly scrambled version. During use, the process runs in reverse: the model starts with a placeholder text full of nonsense tokens and refines it over typically 10 to 100 steps. After each step, the text looks a little more sensible. Throughout this, the model always sees the entire intermediate state — a crucial difference from step-by-step left-to-right generation.

A common misconception is that Diffusion LLMs are necessarily slower than classic models. The opposite can be true: because all positions are processed in parallel, the steps can be effectively accelerated on modern hardware. However, how many steps are actually needed is still a subject of active research.

Diffusion LLMs in products and research

For a long time, diffusion approaches to text were considered an academic niche. This changed significantly in 2024 and 2025. The startup Inception Labs released Mercury, the first commercially available Diffusion LLM, and promoted it as being significantly faster than comparable classic models. At the same time, research kept pace: works such as MDLM and Masked Diffusion showed that diffusion models can match classic models on certain language tasks.

In practice, the term mainly comes up in two contexts: in discussions about the next generation of language model architectures, and in image generation, where diffusion is already the standard. Anyone who reads news about models like Stable Diffusion or DALL·E is thus already halfway familiar with the underlying principle — it is now being applied to language. Whether Diffusion LLMs will ultimately replace or complement classic transformer models in the long run remains an open question. What is clear, however, is that they disprove the assumption that text must necessarily be generated sequentially — that is, one after another.