Definition
Plain language
A way of generating text that mostly fires off a whole block of words at once, then adds a light left-to-right correction so the words agree with each other.
As stated in the literature
A decoding architecture combining a parallel backbone that emits a full block of token probabilities in one pass with a small sequential head that applies per-position corrections conditioned on the previously sampled token; used in DSpark's drafter to get parallel speed with local coherence.
Why it matters: It aims to capture the speed of generating many words at once while still keeping the output coherent, rather than forcing a slow one-word-at-a-time crawl.
For example, it writes a whole sentence in one burst and then makes a quick left-to-right pass to fix words that don't quite fit with their neighbors.
Heard on the show
“That's the semi-autoregressive design.”Episode 179 — How DeepSeek Made One User Faster Without Slowing Down the Crowd