Increasing the number of tokens a model can recognize, typically to support new languages or domains.