I tried designing a model in Blender using surface sub-division and high-resolution vertex colors. What should've taken around 500 kilobytes using a compressed model format ended up taking 15 megabytes when exported to a standard model format as raw polygons.

A combination of adaptive quad-tree detail tessellation and compressed vertex data instead of textures would allow sharing large assets in a less entangled data structure. Each new tessellation step loaded from a compressed model file can let the position and color be stored relative to the average of neighbors. A 3D model for each detail level will be generated as a side-effect from this algorithm.