Step 3 of the leapfrog star model will be officially open-sourced.

On July 31st, Jiaye Xingchen announced that the new generation of their large-scale model Step 3 is officially open source, and has been launched on the Jiaye Xingchen open platform. It is reported that the multimodal capabilities of Step 3 are centered around "lightweight visual pathways" and "stable collaborative training", focusing on solving the token burden and training interference caused by visual input. To achieve this, it adopts a 5B Vision Encoder and uses a double layer 2D convolution to downsample visual features, reducing the number of visual tokens by 1/16, alleviating pressure on context length, and improving inference efficiency.