GitHub radar
How to run Kimi K3 locally on mac: Deltafin
Deltafin allows you to run Moonshot AI’s 2.8 trillion parameter AI model (Kimi K3) locally on your Apple Silicon Mac via an OpenAI compatible API server.
Deltafin is a research project which enables running Moonshot AI’s 2.8 trillion parameter Mixture-of-Experts Language Model (Kimi K3) on a single Apple Silicon Mac/Linux system by streaming model experts from a local disk cache. Kimi’s K2.x family (on which K3 is based) follows a DeepSeek V3 architecture. The research notes explore the inference methodology relative to DeepSeek V4 models. Deltafin also comes with an OpenAI compatible API server which can be used to replace the ChatGPT API for any application. With an M1 Max & 64GB RAM, deltafin produces approximately one token per 15 seconds. Local mode requires around 1.7TB of disk space whereas streaming mode only requires 215GB of disk space (at the expense of slower inference).
Why a vibe-coder should care
Kimi K3 is currently one of the largest public AI models available. Deltafin enables anyone to run this model locally without having to pay any cloud fees or obtain an API key. Although there are significant drawbacks (e.g. you need a powerful Mac & plenty of disk space + time), it shows that cutting edge AI can be run on personal devices right now.
▌ More finds