The multilingual revert risk model supports 44 languages. It has a better accuracy, especially for IP edits, and therefore we would like to use it instead of the language-agnostic model where feasible.
When this model is available in a way that will scale for the volume of activity required for Automoderator we would like to use it. In particular, being able to make requests in parallel would be important - if we have to make requests in series we won't be able to support most large wikis.