High amount of mesh compression.
Vertex positions, normals/tangents and UVs will be quantized to respectively 10, 6, 8 bits per component. Skinning information will be compressed.
See Also: ModelImporter.meshCompression.