So one could take models trained for Marian NMT and make them run on argos? This could be interesting, as it might allow us to re-use already trained models (or use the tools built to train those models).
You could package a Marian model converted for CTranslate2 in an Argos Translate package and it would work. The only issue is that Argoa Translate doesn’t currently support adding a prefix to the source text to tell the multilingual model which language to translate to.
The tradeoff with using a pretrained model from Meta or Microsoft is that it’s larger and more powerful but gives users less control and visibility into how it was trained.
CTranslate2 being able to run third-party models is relatively new so I haven’t implemented anything yet but the plan is to support both FairSeq/Marian models and custom models trained with Argos Train.