TL;DR — Kimi K3 (2.8T-parameter MoE, 1.45 TB of expert weights) running on one M5 Max MacBook Pro with 128 GB, experts streamed from four SSDs. 1.00 tok/s steady decode over a 512-token answer; 1.13 over 128; 0.96 on the public 17-token prompt (upstream reported 0.68). The honest limit: a 512-token prompt takes ~6.3 minutes to its first token. Cause found (prefill re-reads each layer's experts 8×), fix planned, not built. Useful findings: one drive gives ≈52% of four-drive speed, two ≈73%, three ≈90% — the slowest of each layer's 16 reads sets the pace, not total bandwidth. Every number is one cold run with the exact prompt; per-run logs and placement manifests are in k3-public-bench/ . Fork of gavamedia/deltafin (MIT), who built the engine — see CREDITS.md . Instruments: ARGODRIVE .…