Full 2.8T-param Kimi K3 runs on one MacBook at 1 token/sec, streamed off 4 SSDs

Deltafin is an experiment in brute-forcing a frontier model onto consumer hardware. This fork runs Moonshot’s full Kimi K3 — a 2.8-trillion-parameter mixture-of-experts model with roughly 1.45 TB of expert weights — on a single 128 GB M5 Max MacBook Pro by streaming experts from four SSDs on demand. The headline result is a steady decode of about 1 token per second, with the project’s stated point being fidelity rather than speed: nothing is pruned or quantized, and the unmodified K3 target signs off on every token. Small draft models speculate ahead to claw back throughput, but K3 verifies each guess, so output stays bit-for-bit identical to what the full model would produce. This is positioned explicitly against rival efforts that shrink K3’s experts to ~3 bits to fit — faster, but no longer the weights Moonshot shipped, with unmeasured quality costs.