Step-3.7-Flash
StepFunpaired with a 1.88-billion-parameter ViT vision encoder
general family · filed under model architecture
A Vision Transformer (ViT) cuts an image into fixed-size square patches, maps each patch to a vector with one linear layer, adds position information and runs a standard transformer encoder over the patch sequence. Its cost grows with the number of patches, which is the image area divided by the square of the patch size.
Written for the explainer Getting images and video into a decoder: encoders, projectors, and native fusion.
Also called Vision Transformer, ViT vision encoder.
2 quotes from 2 sources, one card per model, each opening with its strongest.
paired with a 1.88-billion-parameter ViT vision encoder
Equipped with a 729M-param Vision Transformer (ViT) featuring hybrid window attention