Cartesia has introduced the H-Nets architecture to replace tokenization

Cartesia has introduced the hierarchical network architecture H-Nets, which completely replaces standard tokenization with end-to-end dynamic chunking.

The developers claim that the new approach provides up to a 10x increase in computational and data efficiency compared to standard Transformer models. The technology also unlocks new neural network behaviors, including contextual tokenization.

The architecture works directly with byte sequences, bypassing the limitations of traditional tokenizers. Detailed technical specifications and test results have been published on the project's official blog.

The new approach reduces raw data processing overhead and could change the standards for building foundational architectures for efficient model training.