Training adjusts model parameters using data and compute over hours or weeks; inference uses the frozen parameters to produce outputs in milliseconds to seconds per request. Training is compute-intensive and typically done on GPU clusters, while inference can run on smaller hardware and must meet latency constraints. The cost profiles differ dramatically: training is a one-time (or periodic) investment, while inference costs scale with usage. Techniques like quantization, distillation, and speculative decoding aim to reduce inference cost without sacrificing quality.