This would also take some load off of the Argos Translate CDN and make model distribution more decentralized in case the CDN went down.
I’m not sure what the Docker/GitHub container repositories allow but models for one language pair are pretty small (~250 MB) and good for people who only want a single language or are trying to get started quickly. We could also build Docker images for common language groups, this would be especially useful if we train more direct (not pivoting through English) models, for example, a direct Spanish to Portuguese model.
All of the language models are currently around 7GB. We could also default/better support LibreTranslate to using a select group of popular models so as Argos Translate adds more languages everyone’s LibreTranslate setup times aren’t increasing.