We present an approach for hierarchical tokenization and reconstruction for resting-state fMRI. Voxel-level BOLD signals are organized into a three-level hierarchy spanning sub-parcel voxel groups, parcels (DiFuMo-1024), and large functional networks (Yeo-17). A shared encoder produces voxel-level representations, which are progressively aggregated to parcel- and network-level tokens via hierarchical mappings. At each scale, tokens are discretized using binary spherical quantization (BSQ) and decoded to reconstruct the corresponding signals. The model is trained using a multi-scale reconstruction objective, encouraging representations to preserve information across voxel, parcel, and network levels. Example reconstructions (bottom) demonstrate recovery of BOLD dynamics at all scales.