It's not a null argument, because LLMs have long moved past "language as in words like english or tamil". Tokens are not units of such language, at least not since visual and audio tokens became first-class objects and there's no actual distinction between them inside the model. That sameness of representation is what it means for a model to be "multimodal".