1\begin{table}[t]2\centering\small3\caption{Energy per generated token on the laptop (M5 Max), OLMoE 4-bit, 256-token greedy decode, held-out general/code prompts; package power from \code{powermetrics} at 1 s; idle baseline 13.2 W. Source: \code{benchmarks/raw/gateG/energy\_laptop/}.}4\label{tab:energy}5\begin{tabular}{llrrrrr}6\toprule7Backend & residency & runs & tok/s & mean W & J/token (gross) & J/token (net of idle) \\8\midrule9mlx-lm & all resident & 12 & 463.0 & 47.2 & 0.101 & 0.073 \\10titan & all resident & 12 & 461.9 & 46.4 & 0.101 & 0.072 \\11titan & 16 slots, cold NVMe, probe & 12 & 42.0 & 18.4 & 0.441 & 0.119 \\12titan & 16 slots, cold NVMe & 12 & 31.9 & 15.9 & 0.498 & 0.086 \\13\bottomrule14\end{tabular}15\end{table}16