A Meta and University of Washington study compares two ways to transfer token-model predictions into byte models. In roughly one-billion-parameter experiments, byte models lagged at low compute but surpassed the compared token models with enough training and matched some performance levels with less data. The claimed maximum advantages over Llama and Gemma are scaling-law extrapolations. This revisits research released in September; longer input sequences, real inference costs and larger-scale reproducibility still need separate evaluation.