Cartesia upgrades H-Net: Enables large models to autonomously learn data segmentation, boosting training efficiency by up to around 10 times.

Insight Beating AI Flash Briefing: Large language models can now learn language independently, but how text is split into tokens is usually determined by pre-trained tokenizers. This saves computing power, yet also constrains how models structure information. Fixed segmentation rules often fall short for unfamiliar words or code; skipping tokenization entirely to process raw bytes directly, however, leads to a sharp rise in computing costs. AI voice firm Cartesia introduced H-Net (Hierarchical Network) in July 2025 to resolve this tradeoff. The framework lets models read raw bytes directly and learn to segment and compress information autonomously. For example, an unfamiliar word may initially split into multiple segments; as it appears repeatedly in context, it gradually merges into fewer segments. This adjustment can occur during inference, with no retraining required. Cartesia has now unveiled a new H-Net version, with notable gains in training efficiency. In experiments with strictly controlled key conditions, computing power needed to hit the same validation loss dropped by ~22%, and computational efficiency for downstream evaluations nearly doubled. In larger-scale trials, the new H-Net matched the average score of open-source model SmolLM3 (trained on ~11 trillion tokens) across six tests using only ~1 trillion token-equivalent data, requiring roughly one-tenth the training compute of SmolLM3. This approach is not limited to text: the team trained H-Net on bacterial and archaeal genomes, finding that when processing human gene fragments, the model could automatically recognize the pattern of grouping every three bases in coding regions. Related DNA modeling experiments saw a 6–7x boost in computational efficiency. Unlike text, speech has no pre-defined word boundaries. Cartesia’s initial H-Net research was actually aimed at enabling voice models to learn phonemes and words directly from continuous audio, without relying on pre-designed segmentation rules.