long term embeddings model hosting solution needed

The BERT-based TF embeddings models some of our classifiers are built on top of were being downloaded from static hosted 3rd party sites. Those URLs were deprecated and they are no longer available. We scrambled and found old copies that we are now hosting on my personal dropbox, which is obviously not a good solution.

This raises a few questions to explore:
1. We need a better intentional approach to saving and running serialized model binaries. How do other people do this?
2. Do we need to retrain the models to match the latest version of TF? is that even possible?
3. Should we move these two embeddings to S3 to join the other models?