LLM streaming latency is judged by time to first token, not total time. Here is the buffer that makes ChatFuse feel fast even when the model is slow.