What are diffusion language models?

A diffusion language model is a model that writes a whole text at once and then corrects it over and over, instead of producing it one word at a time from left to right. The name comes from the mechanism: it starts from an entirely covered sequence, like a page where every word is hidden, and at each pass it uncovers the slots it is most confident about, rereading each time what sits both before and after the point it is working on. These models are usually called DLMs for short, and they are the first alternative that holds at LLM scale to the assumption almost everything else in the field rests on, namely that text is generated in order. The practical difference for a buyer is that the number of passes is a choice: speed becomes a dial to turn, not a fixed property of the model.

See also: Token · The smallest unit of text an AI model processes: not a word, but a word fragment mapped to a numeric id.

How it works, and what changes versus generating in order

A traditional model produces one token at a time and each token depends on the ones already written: it is a chain, and a chain cannot be walked in parallel. That is where the cost of the decode phase comes from, with the graphics card rereading all the weights from memory to produce a single token, its compute units largely unused, and it is the problem that techniques such as speculative decoding work around without removing. A diffusion model removes the assumption: it looks at the whole sequence in both directions and reveals many positions in the same pass, so the work of one step is wide rather than long, and graphics cards are built precisely for wide work. There is a price: positions revealed together are decided without seeing one another, so cutting passes buys time and costs coherence, and quality has to be checked at the configuration you intend to run, not at the one in the demo. The LLaDA paper shows the paradigm holds at LLM scale: an 8-billion-parameter model competitive with LLaMA3 8B, able to surpass GPT-4o at completing a poem in reverse, the task that exposes the reversal curse.

An enterprise example

The documented commercial case is Inception Labs' Mercury Coder, described in a June 2025 technical report: on independent evaluations run by Artificial Analysis the two versions reach 1109 and 737 tokens per second on NVIDIA H100 GPUs, with a claimed speed advantage of up to ten times over speed-optimized competing models and a second place on quality in Copilot Arena. Translated into a decision: for an internal coding assistant, where the person is waiting for the answer and a good draft right now is worth more than perfect prose later, that profile changes the experience noticeably. For a workflow where text is generated overnight and read the next day it changes nothing, and the choice should be made on quality as usual.

Why it matters for decision makers

Three things matter. First, this is still a young family: the supply of models, tools and inference engines is far thinner than the autoregressive one, so adopting it today means accepting a less mature ecosystem in exchange for lower latency. Second, speed here is an explicit choice and belongs in the contract or in the configuration, because nobody notices it by looking at the output of a single request. Third, the security and compliance controls built around text that flows from left to right, from output filtering to partial logs, do not automatically carry over to a model that reveals text out of order: anyone evaluating a DLM in production should budget for a review of those controls, not only of the model.