1\begin{table}[t]2\centering\small3\caption{Qwen3-30B-A3B-Instruct-2507 4-bit on M3U96a (M3 Ultra, 96 GB): mlx-lm vs TITAN all-resident (128 slots, preloaded), 64 tokens, 3 repeats.}4\label{tab:qwen3}5\begin{tabular}{llrrrrrr}6\toprule7Backend & cfg & $n$ & decode tok/s (median) & p95 ITL ms & hit & stall s & TTFT ms \\8\midrule9mlx\_lm & -- & 18 & 121.7 & 8.4 & -- & -- & 50 \\10titan & C=128, none, m=8 & 18 & 121.6 & 8.5 & 1.000 & 0.00 & 51 \\11\bottomrule12\end{tabular}13\begin{tablenotes}\footnotesize14\item Sources: \code{benchmarks/raw/gateC/qwen3\_m3u96a/baseline.jsonl}.15\end{tablenotes}16\end{table}17