Hi Sarah from Open Privacy / Cwtch team here - the main major difference is that Cwtch servers are completely untrusted under the risk model - they don't learn anything about the groups they are hosting, who is a member of which group, or who each message is for.

The design for groups is still in flux, and they are marked experimental but there are a few more details in our Secure Development Handbook https://docs.openprivacy.ca/cwtch-security-handbook/groups.h...

Metadata resistant group comms is still a fairly large open research problem, and we are also working on the research side to reduce some of the bandwidth requirements that are currently required by our group protocol: https://git.openprivacy.ca/openprivacy/niwl