Kimi K3 runs at ~1 token/s on MacBook Pro with 4 SSDs
A report shows Kimi K3 (2.8T parameters, 1.45 TB of expert weights) running on an M5 Max MacBook Pro (128 GB) with experts streamed from four SSDs; a 512-token answer achieved ~1.00 token/s steady decode. Measurement date is 2026-09-08. Related projects mentioned are Deltafin and ARGODRIVE.