diffusion language model
A language model that drafts a whole answer roughly and then refines many words in parallel, instead of writing one token after another.
Nearly every chat model you have used writes left to right: one token, then the next, each one conditioned on everything before it. That is called autoregressive generation, and it puts a hard ceiling on speed, because token 200 cannot be written before token 199 exists.
A diffusion language model borrows the trick from image generators. It starts with a rough, noisy version of the entire answer and sharpens it over several passes, correcting many positions at once. Because the work happens in parallel, the same graphics card produces far more text per second. The historical tradeoff was quality on long chains of careful reasoning, since the model commits to a shape early, and closing that gap is what recent releases like Mercury 2.5 are trying to prove.