Open-source fast LLM serving framework from LMSYS — 5x faster than vLLM on structured generation. The bleeding-edge inference engine for agents.