Contextual embedding
Also: contextualized embedding, ELMo
Instead of one fixed vector per word, a deep model (ELMo with LSTMs, BERT and GPT with transformers) produces a vector for each token occurrence that depends on its neighbors. This resolves polysemy, which static word embeddings cannot.
Example
"bank" in "The bank is by the river" and in "I need to go to the bank" share one static word2vec vector, but BERT gives them two distinct vectors, one near "shore" and one near "money".