The 3D models are compressed with 'Draco' compression. So for example the largest asset is a model of the entire world which is 333 kB. It consists of 81k vertices. Unoptimized that would mean at least 3 32 bit floats per vertex, which would be at least 972 kB uncompressed. So Draco compression is doing a pretty good job getting the model sizes down.
I'd have to dig into it more but I suspect if they wanted to they could have trimmed the size down quite a bit, but it may not be worth the effort. They could have used more repeat 3D models, but instead it looks like they went for a more artist driven look where most of the world is uniquely modeled.
It also looks like they're sending a bunch of lower res levels of detail per model over the network. That also is a tradeoff. Still with network speeds as they are nowadays it may be faster / simpler to send those over the wire each time instead of regenerating the level of detail.
The textures are a similar story. Well compressed but they're sending procedural noise over the network. Those could be generated but it's also trivial to send them over the network.