There's Russian research on this, they swapped the tokenizer of an existing model (IIRC something from the Qwen family) for one which had far more Cyrillic tokens. After some post-training, the model kept its original quality, but it used way fewer tokens.