Cartesia upgrade H-Net, claiming that training efficiency improved by up to approximately 10 times

AT A GLANCE

Cartesia releases a new version H-Net, reportedly capable of autonomously segmenting raw bytes and significantly reducing training compute in some experiments.

Article

Cartesia Upgrade H-Net: enabling large models to learn data segmentation themselves, improving training efficiency by up to approximately 10 times

Dongcha Beating AI News: Large models can already learn language on their own, but how text should be split into Token is usually still determined by a tokenizer trained in advance. This saves computing power, but also limits how models organize information. When encountering unfamiliar vocabulary or code, fixed segmentation rules may not be suitable; if no tokenization is used and processing is performed directly byte by byte, computational costs increase substantially.AI Speech company Cartesia in 2025 year July proposed H-Net(a hierarchical network) to address this trade-off. It enables models to read raw bytes directly and learn how to segment and compress information themselves. For example, when encountering an unfamiliar word, it may initially be split into multiple segments; as the word repeatedly appears in context, it is gradually merged into fewer segments. This change can also occur during inference, without requiring retraining.

Cartesia has now released a new version of H-Net, with a clear improvement in training efficiency. In experiments with key conditions strictly controlled, the computing power required to reach the same validation loss was reduced by approximately 22%, while computational efficiency on downstream evaluations approximately doubled. In larger-scale experiments, the new version of H-Net used only approximately 1000000000000 Token of equivalent data to match the average score across six tests of an open-source model trained on approximately 11000000000000 Token's open-source model SmolLM3. The training compute required to achieve this result was approximately one-tenth of that of the latter.

This method is not limited to text. The team trained H-Net using bacterial and archaeal genomes and found that, when processing human gene fragments, the model could independently identify the pattern of coding regions being grouped every nucleotides. The computational efficiency of the related DNA modeling experiments improved by approximately 6–7 times.

Unlike text, sound does not have ready-made word boundaries. Cartesia's initial research into H-Net was in fact intended to enable speech models to learn phonemes and words directly from continuous sound, without relying on pre-designed segmentation rules.

Original link https://m.theblockbeats.info/flash/374378

Key points

The report refers to H-Net as a “hierarchical network” and states that Cartesia proposed this method in 2025 year July.
H-Net It directly reads raw bytes, allowing the model to learn how to segment and compress information.
The report cites an example in which, after unfamiliar words repeatedly appeared in context,H-Net they could be merged into fewer segments, and this change can occur during inference without retraining.
In experiments with key conditions strictly controlled, the new version described in the report H-Net required approximately 22% less compute to reach the same validation loss, while computational efficiency on downstream evaluations approximately doubled.
In larger-scale experiments, the reported new version H-Net uses approximately 1000000000000 Token equivalent data, matching the average score across six tests achieved by training on approximately 11000000000000 Token of the SmolLM3, with the required training compute being approximately one-tenth that of the latter.
In DNA modeling experiments,H-Net when processing human gene fragments, it identified the pattern of coding regions being grouped in sets of bases, with the computational efficiency improvement described in the report being approximately 6–7 times.
The report states that Cartesia initially researched H-Net to enable speech models to learn phonemes and words from continuous sound without relying on predesigned segmentation rules.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Traditional large language models typically first use a fixed tokenizer to split text into Token。H-Net instead tries to receive raw bytes directly, allowing the model to determine on its own, based on the data, how to group and compress them. The report suggests that this approach may combine flexible segmentation with computational efficiency and can be used for text,DNA and audio.

Why it matters to readers

If the experimental results reported can be reproduced, models may reduce their reliance on fixed tokenization rules and process different types of data with less compute.

Beginners can understand it as follows: the model no longer relies entirely on a manually predefined “word-segmentation method,” but instead learns how to group data while processing it; however, the efficiency figures in the report still need to be verified against the original experiments.

Risks and unknowns

The evidence is a media brief quoting secondary reporting and does not provide the new version’s H-Net original paper, code, or complete experimental setup.
Whether the efficiency results of approximately 22%, approximately 2 times, approximately 6–7 times, and approximately 10 times can be replicated by independent teams cannot be confirmed from the available evidence.
It remains unclear whether the advantages observed in controlled experiments can extend to practical speech models and larger-scale language models.
H-Net and SmolLM3 Whether the specific model scale, data composition, and evaluation conditions are fully comparable is not specified by the evidence.

Cartesia upgrade H-Net, claiming that training efficiency improved by up to approximately 10 times

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center