Radar ·
H-Nets: byte-level hierarchical networks for language modeling
A new version of H-Nets, a tokenizer-free architecture that learns from raw bytes. The post reports 3-5x efficiency gains and parity with an open-source model, as author claims not yet independently verified.
According to the post dated October 28, 2026, the author announces a new version of H-Nets, hierarchical networks that learn end to end from raw bytes, without a tokenizer. The stated numbers are 3-5x efficiency gains in scaling law experiments and parity with an open-source model trained on 11T tokens, using about 10x less data. The available excerpt is truncated, so it gives no details on methodology, benchmarks or experiment setup.
These results are the author's claims and must be verified in the full text, the original papers and code, and through independent review. To consult them, open the original post from your bookmark reader, look for the methodology and compare the numbers with the cited sources. To study the material with AI, use only public excerpts: do not paste personal data, internal documents or credentials, and replace names or identifiers with markers before sending. Rota Nacional detects CPF, CNPJ, e-mail, phone numbers and names and applies the organization's policy, but this protection does not replace the user's own review.