Parallax

by Gradient · Apache-2.0 · 1,420 GitHub stars · Quiet · v0.1.2, Dec 2025

A scheduler that places layer slices of a supported model across Macs (MLX) and NVIDIA PCs (SGLang or vLLM), on a LAN or across the internet.

Mac: Yes
NVIDIA PC: Yes
AMD PC: No
Linux, processor only: No
Raspberry Pi and boards: No
Windows: Partly
Android: No
iPhone, iPad: No

Go to Parallax Source code

At a glance

Best for: A mix of NVIDIA PCs and Macs, possibly on different networks, serving one of the models on its list.
Runs on: Macs through MLX and NVIDIA cards through SGLang or vLLM on Linux; Windows means WSL2 with an NVIDIA card.
How it splits a model: Pipeline across machines, with the scheduler choosing each machine's layers and the fastest route for each request.
Connections: Its own peer-to-peer layer over TCP and UDP, with NAT traversal; no Thunderbolt or RDMA.
How devices join: A scheduler machine; the others join it.
Models: Hugging Face weights on NVIDIA machines and a matching MLX build on Macs, for about 40 listed models.
API: OpenAI-style chat endpoint.
Rules: One machine runs the scheduler (parallax run), the others join (parallax join).

Get started

macOS

git clone https://github.com/GradientHQ/parallax && cd parallax && ./install.sh

Linux

docker pull gradientservice/parallax:latest

Windows

Parallax_Win_Setup.exe (sets up WSL2 and CUDA)

Then:

parallax run   # on the scheduler
parallax join   # on each other machine

Limits to know

No changes since July 2026 and no release since December 2025.
Pipeline only across machines.
Remote mode shares machine addresses with a public relay.

Open issues worth reading:

Fails across networks with symmetric NATparallax issue #197 (opens in a new tab)

Sources:

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.