Bolmo’s architecture unlocks efficient byte‑level LM training without sacrificing quality - Tech Sentiments