Benchmarked the MacBook M4 Pro (24 GB unified, Metal 4) against the iMac (Radeon Pro 575, Vulkan). The E2B Q4 hits 90 tok/s on the M4 — that's instantaneous, faster than reading speed. The E4B at Q8 does 42 tok/s. Both feel like native apps. But the 31B? The M4 Pro loads it into unified memory just fine — no swap, no thrashing — and runs at 12.4 tok/s. Usable for batch processing, too slow for conversation. The iMac's Vulkan path actually beats it at 7.5 tok/s because it's offloading to the discrete GPU's 4 GB VRAM while the CPU handles the rest. Different architectures, different tradeoffs. The M4 is the daily driver. The iMac is the overnight workhorse.