Mac Studio M2 Ultra Qwen3-235B-A22B 4-bit decode, MLX

THIRD-PARTY MEASURED (Silicon Score table; MLX runtime on M2 Ultra unified memory). 4-bit quant, MoE - only ~22B active params read per token, which is why the big-MoE number far exceeds the ~9-11 t/s dense-70B estimate on the same ~800 GB/s part.

192GB config; the capacity play for 70B at 8-bit + long ctx or 235B-class MoE. Used ~$3,200-4,200 Oct-2026.