These model's tokenizer did not use HF's fast tokenizer, resulting in variations in how pre-tokenization was applied. Resolved in latest versions.