That's a great question. Our current setup is p2p WebRTC and should scale to many small rooms (up to ~12 people) fairly well. That said, there is a TURN server for NAT traversal and a signal server for coordination.
We'll try to get a better picture of where the limits are this week. I don't have a good answer atm to be honest.
Assuming that each peer is connected to every other peer via a mesh network [see this image for reference: https://github.com/feross/simple-peer/blob/master/img/full-m...], each outgoing stream (esp. audio / video) is likely going to be duplicated, per recipient.
Scalability over a mesh network is fully dependent on CPU and network performance of all of the connected devices, and I'd doubt it could handle 12 participants if there is video involved, unless all participants are running relatively high-end and modern devices, with optimal network conditions.
You'll need a SFU or an MFU running on the server to handle larger rooms, while enabling all connected devices to only have to send one output stream per media type, regardless of how many connected participants there are.