Research & academia
Now in Nature: Retrofitting language models to operate over bytes
Bolmo’s byteifying process turns capable subword models into byte-level models with a relatively short additional training run—and the approach now extends beyond Olmo.
Why it matters: For researchers exploring byte-level models, byteifying offers a way to build on capable subword models rather than training from scratch. The reported results are near-matches, not uniform superiority, and applications to images or audio remain prospective.