Cake

by Simone Margaritelli (evilsocket) · FAIR License (source-available, no business use) · 3,133 GitHub stars · Quiet · no releases

A Rust project that splits a model across phones, tablets, Macs and PCs; the main machine streams each one the layers it needs. Quiet since April 2026.

Mac: Yes
NVIDIA PC: Yes
AMD PC: Partly
Linux, processor only: Yes
Raspberry Pi and boards: Yes
Windows: Partly
Android: Partly
iPhone, iPad: Yes

Go to Cake Source code

At a glance

Best for: Hobby experiments that include iPhones and iPads, for non-commercial use.
Runs on: Metal, CUDA and Vulkan builds, an Android CPU worker and an iOS worker app; built from source.
How it splits a model: Pipeline.
Connections: TCP; data isn't encrypted, so use a VPN on untrusted networks.
How devices join: Automatic with a shared cluster key.
Models: Hugging Face safetensors, or GGUF converted to full precision at load.
API: OpenAI compatible.
Rules: Workers need no model files; the main machine sends them.

Get started

macOS

cargo build --release --features metal

Limits to know

No commits since April 2026 and no releases.
Not open source: its licence forbids business use.
No published multi-device benchmarks.

We don’t estimate speeds for Cake.

Sources:

evilsocket/cake on GitHub

All clustering tools

Estimates come from published hardware specs and our open formulas, not from running the model on your machine. Real speed varies with runtime, drivers and settings. Numbers marked “Measured” come from a linked benchmark.