Cartesia's H-Nets Learn Language From Raw Bytes, Cutting Compute 10x

Scaling Hierarchical Networks

Cartesia scales its hierarchical networks (H-Nets), which replace hand-designed tokenizers with learned, end-to-end chunking of raw bytes. A controlled scaling ladder shows a one-stage H-Net gains 3-5x compute efficiency purely from learned tokenization, while a larger two-stage H-Net beats a matched open-source model using 10x less compute and data. H-Nets also show in-context tokenization and promise across biology and vision.