Small, tokenizer-free language models trained entirely from scratch. No BPE, no subword vocabulary — these models read and write raw bytes or characte