MLX distributed

by Apple (ml-explore) · MIT · 29,079 GitHub stars · Active · mlx v0.32.4, mlx-lm v0.32.2, Oct 2026

Apple's own machine-learning framework can spread one model across several Macs with its launcher, and is what EXO builds on; you set it up from the command line.

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

Go to MLX distributed Source code

At a glance

Best for: Mac owners comfortable with a terminal who want the most direct, lowest-overhead way to run one model on several Macs, especially Mac Studios on Thunderbolt 5.
Runs on: Built for Apple silicon. Linux builds exist for NVIDIA cards (with NCCL between machines) and for processors only; Macs and Linux machines aren't mixed.
How it splits a model: Tensor or pipeline, per model: about two dozen model families support tensor split and about fifteen pipeline.
Connections: A TCP ring over Ethernet or Thunderbolt, or JACCL: RDMA over Thunderbolt 5 (macOS 26.2+), which Apple calls an order of magnitude lower in latency and needed for tensor split.
How devices join: Manual: a host file and passwordless ssh between the Macs (mlx.distributed_config can write it).
Models: MLX models from Hugging Face.
API: OpenAI compatible (mlx_lm.server), plus chat and generate commands.
Rules: Same Python and paths on every Mac. With tensor split each Mac needs every weight file; pipeline mode downloads only its own share.

Get started

macOS

pip install mlx mlx-lm

Linux

pip install mlx[cuda13] mlx-lm

Then:

mlx.distributed_config --hosts mac1,mac2 --over thunderbolt --backend jaccl --auto-setup --output hosts.json
mlx.launch --backend jaccl --hostfile hosts.json -- python -m mlx_lm chat --model mlx-community/<model>

Published speeds

Devstral 2 123B 4-bit on 1, 2 and 4 Mac Studios (M3 Ultra) over RDMA: 9.0, 15.8 and 26.2 tok/s (Apple maintainer's benchmark). Source
GLM 4.7 358B 6-bit, same Macs: 16.9, 22.3 and 27.7 tok/s. Mixture-of-experts models gain less from the tensor split than dense ones. Source

Measured by the project or its users, for one person chatting. Our own estimates for any set of devices are in the cluster builder.

Limits to know

No app or discovery: host files and ssh.
Its server is for personal use, not production.
Community reports describe an early RDMA driver with occasional hangs.

Open issues worth reading:

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.