Low amount of mesh compression.
Vertex positions, normals/tangents and UVs will be quantized to respectively 20, 8, 16 bits per component. Skinning information will be compressed.
See Also: ModelImporter.meshCompression.