Do all these models have any significant architectural differences or training data sources? What are the factors going into the diversity of their performance?

The article posted is basically entirely about that.