Importantly, though, LLMs do not take the embeddings as input during training; they take the tokens and learn the embeddings as part of the training.

Specifically all Transformer-based models; older models used things like word2vec or elmo, but all current LLMs train their embeddings from scratch.