2 min read

How I Supercharged an Underpowered Mac Mini to Run a Local 30B-Parameter Model

Source: http://x.com/i/article/2041159743608152064

I bought a Mac Mini in early February, like a lot of people, as a low-cost way to dive into

@openclaw and local AI tools. For a while, it worked great for my setup. It handled OpenClaw well and was perfectly fine for smaller models, but once I started experimenting with larger local models, I realized pretty quickly that 16GB was not enough. After seeing a few posts on X about clustering Apple Silicon devices, I picked up some Thunderbolt cables and decided to test the approach myself. I initially assumed that daisy-chaining devices would simply pool VRAM and let me run any model directly through LM Studio or Ollama, but that turned out to be dead wrong.

https://x.com/aibra/status/2040615048653975581

This morning I successfully ran Qwen3-30B-A3B-4bit (19GB) on my cluster using

@exolabs. Getting there took some real trial and error. If you’re considering this setup, here’s the workflow that finally worked for me.

How I got it working 1.

Enable RDMA on each device:

Power off the Mac, hold the power button to enter Recovery, then open Utilities → Terminal and run: rdma_ctl enable. Press Enter and let the device reboot.

Install Exo on each device:

Download

https://assets.exolabs.net/EXO-latest.dmg

. Let Exo launch at startup, then open the dashboard.

Connect devices over Thunderbolt:

Use Thunderbolt 5 cables between devices. On a base Mac Mini, use the port marked with the lightning icon. Exo will establish a network over the wired connection.

Download the model on each device:

Exo handles this through the dashboard when you select a model.

Load the sharded model:

Exo splits the model into shards across devices. In my case, one shard landed on the M4 Mac Mini (16GB) and the other on an M4 MacBook Pro (24GB).

Running a model on the Cluster

Model launch kept failing with no clear errors in the @exolabs UI. In my setup, the default RDMA pathway was failing to load the model. I suspect hardware/path compatibility in my specific mix of devices, but I could not confirm root cause.

I used Codex to troubleshoot, and it identified a working path: switch from RDMA to MLXRing. Since that toggle is not currently exposed in the Exo UI, Codex generated a Python script to recreate the instance with MLXRing. I also wrapped that into an OpenClaw skill so I can trigger it from Telegram instead of using Terminal each time.

Link to script repo:

exo-mlxring-loader

Using the cluster as a local model endpoint

Once the model is loaded, Exo exposes an OpenAI-compatible API at: http://localhost:52415/v1

That means you can plug the cluster into tools like

@openclaw or @NousResearch Hermes by pointing your agent/model config to that endpoint.

Final Notes and Thoughts:

The performance boost was not dramatic, but this setup unlocked models I could not run before. Biggest practical win was capacity, not raw speed. Performance improved when I closed non-essential apps on both machines. Power draw on the MBP increased to about 40W under load (from roughly 3W idle), and temps rose, but stayed within a reasonable range.