Nature: 'byteification' converts subword LMs to byte-level

A new Nature paper published 7 October 2026 says you can convert an existing subword language model into a byte-level one for less than 1% of a typical pretraining budget. The authors call the recipe 'byteification.' Total training cost in the paper: 49.1B tokens, split into a 9.8B-token first stage and a 39.3B-token second stage.

The headline result is a 7B model the authors name Bolmo 7B. Against BLT 7B, the paper's byte-level baseline, Bolmo 7B posts a '+16.5% absolute improvement in STEM tasks.' The abstract frames the broader claim this way: the method lets converted models 'excel on character-level reasoning tasks, achieving practical inference speeds' and offers 'advantages in domains requiring fine-grained textual understanding.'

The authors go further on the significance line. Their results, they write, 'remove a long-standing performance barrier to end-to-end byte-level language modelling, demonstrating that models operating on raw text encodings can scale competitively.' The author list spans AI2, Cambridge, University of Washington and Edinburgh, including Luke Zettlemoyer and Noah A. Smith; two researchers in our Who's Who directory shared the link on the day it dropped.

The abstract does not publish a latency number for 'practical inference speeds,' name the specific STEM benchmarks that go into the +16.5%, or compare Bolmo 7B against a modern subword 7B. The claim being made is parity-approaching, not parity-on-the-table.