Stratumus Journal Research · How it works
Anthropic’s researchers recently published a study of AI models that might one day copy themselves onto cheap internet devices. To show the hardware side was possible, they split a 26-billion-parameter model across 15 off-brand Android TV boxes, each box holding two of the model’s layers. It ran at 0.39 tokens per second. (Anthropic)
Mac Pool uses the same basic technique, splitting a model by layers across machines, for the opposite reason: to run a model bigger than any one of your Macs, on Macs you own and added yourself. Here is how it works, what it costs over Wi-Fi, how it is locked down, and what we learned.
Why split a model at all? #
A model’s weights have to sit in memory while it answers. Qwen3 32B at 4-bit (Q4_K_M) is about 20 GB of weights, before any room for the conversation. That is more than a 16 GB or a 24 GB Mac can hand to its GPU. Together, those two Macs have 40 GB.
So the point of Mac Pool is memory, not speed. Splitting a model that already fits on one Mac does not make it faster. Splitting a model that doesn’t fit makes it possible.
Split by layers, not inside them #
There are two common ways to spread one model over several machines:
- Splitting inside each layer (tensor parallelism). Every machine works on every layer, so they talk once per layer for every token. Over Thunderbolt 5 that can make a model faster;exo does this well. Over Wi-Fi, dozens of round trips per token are too slow.
- Splitting by layers (pipeline parallelism). Each machine holds a block of whole layers. For every token, only the activations at the boundary between blocks cross the network: one vector the width of the model, tens of kilobytes. That is why the TV boxes, and Mac Pool, use it.
The Mac that holds the conversation (we call it the head) gives every other Mac a share of the layers it has room for, and leaves room on each one for macOS and your apps. Stratumus for Mac sets all of it up for you, so nobody opens Terminal. Mac Pool builds on the open-source llama.cpp project.
What it costs over Wi-Fi #
These are measurements from our first real run, on October 3, 2026: a Mac mini (M4, 16 GB) and a MacBook Pro (24 GB) on ordinary home Wi-Fi.
| Round trip between the Macs | about 12 ms (up to about 22 ms) |
|---|---|
| Bandwidth between the Macs | 138 to 183 Mbps |
| Cost of encrypting the link | about 0.08 ms per message |
| First load of a share of weights | about 17 MB/s over the network |
| Re a 32B model later | about a minute; each Mac reads its share from its own disk |
| First reply, DeepSeek-R1 32B on both Macs | about 100 s, most of it the model’s own thinking |
The network is not the bottleneck it sounds like. Moving the weights is slow once, and after that each Mac keeps its share on disk, so only the small activations travel. We have not yet published pooled tokens per second or time to first word, and the about 100 seconds above is still our only published figure. A newer measured run will be published here, with its method.
How it is locked down #
The study’s example spread through devices with an open port that anyone could talk to. Mac Pool never leaves one open:
- The part that runs the model listens only inside its own Mac. Nothing on the network can reach it directly.
- Macs reach each other only through an encrypted link , keyed from your Mac Pool key. A device without the key can’t connect.
- The link listens only on a private home-network address and refuses public ones. It forwards to the worker and nothing else.
- A Mac joins only when you add it from your signed-in account. Being on the same network is never enough. Removing a Mac creates a new key.
One thing we want to improve: Stratumus creates your Mac Pool key on its server and stores it encrypted, so today it is not end-to-end. The model traffic between your Macs stays on your home network and never passes through us.
What we learned #
The first time DeepSeek-R1 32B ran across both Macs, it took several fixes before it answered. The lesson surprised us. On Apple silicon, with memory shared between CPU and GPU, the hard part of splitting a model is not the network. It is counting memory honestly. Much of the work in Mac Pool goes into making sure every Mac has room before a model loads, so you never have to think about it.
Why Macs, and why only yours #
The TV boxes in the study had 4 GB of memory each and slow memory buses, which is why 15 of them managed 0.39 tokens per second. One 16 GB Mac can hold as many layers as four of them, and Apple silicon moves data through memory many times faster. That is the engineering reason Stratumus starts with Macs.
The other reason is consent. Mac Pool only ever uses Macs you own and added yourself. It does not borrow other people’s devices, smart TVs or spare computers on the internet. Useful intelligence. Human control.