⏱ 7 min read How mixture-of-experts models split parameter count from compute per token, and the routing, memory, and serving tradeoffs interviewers test. Read article