Apple

Apple has released a suite of vision encoders under a permissive license, surpassing both CLIP and SigLIP. Instead of using both a text and a vision encoder which are then trained to maximize the similarity of the representations, Apple uses a single vision encoder, followed by a shared multimodal decoder, which processes both the image patches and the text in sequence

Apple released a baseline for the open collaboration DataComp for Language Models (DCLM). This is a stronger model than OpenELM was, so let's see where they take this.