Stratumus – Qwen3 32B on a 16 GB and a 24 GB Mac over home Wi-Fi
- Source
- stratumus.com
- Author
- jakebowser
- Date

Context
Stratumus's Mac Pool runs one AI model across several Macs you own so that a model too big for any single Mac's memory can run at all. Its purpose is memory, not speed. The founder's write-up uses Qwen3 32B, an open model from Alibaba's Qwen team, as the sizing example: compressed to about four bits per weight, its weights take roughly 20 GB, more than a 16 GB or 24 GB Mac can hand to its GPU, while the two machines together have 40 GB.
The split is by layers. Each Mac holds a block of whole model layers sized to its free memory, and for every generated word piece only the intermediate result at the boundary between blocks crosses the network, a single vector of tens of kilobytes. The alternative, splitting the work inside every layer, needs a network round trip per layer per ; Stratumus says that can pay off over Thunderbolt 5, as exo does, but is too slow over Wi-Fi. Mac Pool is built on the open-source llama.cpp runtime.
The published evidence is one run on October 3, 2026, between an M4 Mac mini with 16 GB and a 24 GB MacBook Pro on home Wi-Fi. The company reports about 12 ms between the machines, 138 to 183 Mbps of bandwidth, a slow first copy of the weights at about 17 MB/s, and reloads of about a minute once each Mac keeps its share on disk. Its only end-to-end figure is about 100 seconds to the first reply from DeepSeek-R1 32B, most of it the model's own reasoning. No tokens-per-second number has been published, so the Qwen3 32B in the headline is a sizing example rather than a measured result. Stratumus says the hard engineering problem was counting memory accurately on Apple silicon, where the CPU and GPU share it, not the network.
The post frames Mac Pool against Anthropic's research on small models that might copy themselves onto cheap internet devices (earlier Intel coverage), which, as Stratumus describes it, split a 26-billion-parameter model across 15 Android TV boxes at 0.39 tokens per second. Stratumus says its version only uses Macs the owner adds from a signed-in account, keeps the model process unreachable from the network, and joins machines through an encrypted link that refuses public addresses. It also states one gap itself: the pool key is created and stored encrypted on Stratumus's server, so the setup is not end-to-end encrypted, although model traffic stays on the home network.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
- quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
Checking sign-in…
Loading comments…

