For local serving, throughput is often limited by memory bandwidth during decoding and by compute during prefill. Parameter count, quantization, context length, and concurrency all move it. Measure tokens per second per sequence and total throughput separately; they answer different questions. See time to first token.