According to CoinMeta, Cartesia has announced an upgrade to H-Net, aimed at enabling large models to learn autonomously how to split data, resulting in a nearly 10-fold increase in training efficiency. This technology allows models to directly read raw bytes and learn information segmentation and compression methods on their own, performing particularly well when dealing with unfamiliar vocabulary. The new version of H-Net has shown in experiments that it requires about 22% less computing power to achieve the same verification loss, and the computational efficiency in downstream evaluations has nearly doubled. In larger-scale experiments, H-Net used only about 1 trillion token equivalent units of data to match the performance of the open-source model smollm3 which trained with about 11 trillion token units of data in six tests, requiring only one-tenth of the latter's computing power. In addition, this method is also applicable to genomic data, with a computational efficiency improvement of about 6-7 times.