Per https://meta.wikimedia.org/wiki/Machine_learning_models/Proposed/Multilingual_revert_risk:

This model was developed to improve the performance of it's Language Agnostic (RRLA) version. The Multilingual version shows a better performance, especially for IP edits

so we should probably use this as the default, for the 47 languages it supports, and switch to RRLA if the language isn't supported.

Conceptually, it would probably make sense to group both models in the same model class in ORES, but I am not 100% sure about that.