Open Models Almanac
Methodsmodel architecturemultimodal architecture

specific method · filed under model architecture

Hierarchical patch encoder

Inkling's "hierarchical patch encoder" is a four-layer MLP stem (hMLP) that folds each 40×40-pixel patch into one embedding in stages of growing size, with no attention between patches. It is the reason the lab can call the model encoder-free while still naming an encoder.

Written for the explainer Getting images and video into a decoder: encoders, projectors, and native fusion.

Also called Hierarchical patch encoder for images.

sources
4
model
1
lab adopt it
1
strongest
core

Documented in

Thinking Machines Lab

Evidence

4 quotes from 4 sources, one card per model, each opening with its strongest.

How sources treat it
core 4

Inkling

Thinking Machines Labadopted4 sources

images and video are encoded via a hierarchical patch encoder

coremodel architecturein InklingThinking Machines Lab
3 more from this model

images and video are encoded via a hierarchical patch encoder

coremodel architecturein InklingThinking Machines Lab

Images and videos are encoded via a hierarchical patch encoder

coremodel architecturein InklingThinking Machines Lab

images are encoded via a hierarchical patch encoder

coremodel architecturein InklingThinking Machines Lab
↑