Open Models Almanac
Methodsmodel architectureprediction head

single source · model architecture

Recursively shared Multi-Token Prediction head

Uses a single recursively shared MTP head, initialized from the first head while discarding a second head.

The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See prediction head for the methods filed beside it.

Also called single, recursively shared head, recursively shared MTP formulation.

Filed under model architecture › prediction head