They didn't train the vision encoder either, it's unchanged SigLIP by Google.