A new generation of Hierarchical Networks (H-Nets) can learn directly from raw bytes rather than tokenized data. They match the performance of an open-source 3B-parameter Transformer while using 10x less compute and training data.