Byte-level H-Net reportedly reaches SmolLM3's six-task average with roughly one-tenth the training tokens

Cartesia reports that its updated byte-level H-Net reached SmolLM3's final average across six benchmark tasks after roughly 1 trillion training tokens, versus roughly 11 trillion for SmolLM3, with proportionately less training compute. The model learns to form chunks from bytes instead of using a fixed tokenizer. The comparison uses different model sizes and training data, which Cartesia cautions make a clean match difficult.