Attention is shared. It's ~30% of params here. So ~2B params are shared between experts and ~5B params are unique to each expert.