>>105750446
They have separate vision experts. They mention in the technical report that the visual experts contain one-third the parameters of textual experts.