Cartesia upgrades H-Net: enabling large models to learn data segmentation on their own, boosting training efficiency by up to about 10 times

Beating AI News Flash: Large models can already learn language on their own, but how text should be split into tokens is usually still determined by a pre-trained tokenizer. This saves compute, but it also limits how the model organizes information. When encountering unfamiliar words or code, fixed splitting rules may not be appropriate; if there is no tokenization at all and processing is done directly byte by byte, computational costs increase substantially.

AI voice company Cartesia proposed H-Net (Hierarchical Network) in July 2025 precisely to solve this contradiction. It lets the model read raw bytes directly and learn on its own how to split and compress information. For example, when encountering an unfamiliar word, it may initially be split into multiple segments; as the word repeatedly appears in context, these segments gradually merge into fewer ones. This change can also occur during inference without the need for retraining.

Now Cartesia has announced a new version of H-Net, with a clear improvement in training efficiency. In experiments with strict control of major conditions, the compute required to reach the same validation loss was reduced by about 22%, and computational efficiency in downstream evaluations roughly doubled. In a larger-scale experiment, the new version of H-Net, using only about 1 trillion token-equivalent data, matched the average score of the open-source model SmolLM3 trained on about 11 trillion tokens across six tests, while requiring only about one-tenth of the latter's training compute to achieve this result.

This approach is not limited to text. The team trained H-Net on bacterial and archaeal genomes and found that when processing human gene fragments, the model could identify on its own the pattern of every three bases in coding regions. The computational efficiency of related DNA modeling experiments improved by about 6–7x.

Sound, unlike text, does not have ready-made word boundaries. Cartesia's initial research on H-Net was actually intended to enable speech models to learn phonemes and words directly from continuous audio without relying on pre-designed splitting rules.