Intel / repo

Kimi K3 (2.8T) at 1 token/s on a MacBook Pro, streamed from four SSDs

Source: github.com
Author: Argonautlabs
Date: Published Sep 8, 2026

Context

Deltafin demonstrates a way to run Moonshot AI's Kimi K3 on a laptop by its expert weights from four SSDs. The project's benchmark report describes an M5 Max MacBook Pro with 128 GB of memory, reaching about one output per second during a 512-token answer. That figure is the steady rate after the first token: a 512-token prompt took approximately 6.3 minutes before any output began, which the report attributes to prompt processing. The experiment therefore shows how storage can accommodate a model larger than memory, while making the waiting time visible alongside the headline generation rate.

Terms in this piece · Glossary

streaming — Sending a model's response token by token as it is generated, so the reader sees text immediately instead of waiting for the whole answer.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.

Read the source github.com