llama.cpp and vLLM have dominated local inference over the past few years, with their core advantage being universality: supporting a wide range of model architectures, covering multiple hardware backends, and having an active community.