Bolmo 7B: model reads bytes instead of tokens, conversion cost less than 1% of pretraining

Ai2 researchers led by Benjamin Minixhofer converted existing open models to byte-level using 49.1 billion tokens, less than 1% of the pretraining budget. The paper on the method was published in Nature on Oct 7, 2026, and the preprint has been on arXiv since Dec 17, 2025.