Running a 397B model on a 64 GB Mac
- Outlier Plus is a 397B-parameter mixture-of-experts, 17B active per token. The weights are a 209 GB download.
- It runs on a 64 GB Mac at roughly 10 GB peak resident memory, because only the experts a token needs are ever in memory.
- Measured speed: 1.59 tokens per second on an M1 Ultra. That is slow enough to matter and fast enough to use for hard questions.
- The constraint is disk, not memory. You need room for 209 GB of weights.
The usual rule for local models is that the weights have to fit in memory. A 209 GB model therefore needs a machine nobody has on a desk. Mixture-of-experts architectures break that rule, and our paged engine is built around the gap.
Why it fits
In a mixture-of-experts model most parameters are idle for any given token. Plus routes each token to 10 of 512 experts, so about 17B parameters do the work while the other 380B sit unused. Those unused experts do not need to be in memory. They need to be reachable.
Our V9 paged engine keeps the always-active parts resident and streams the rest from SSD as routing demands them. Peak resident memory measured 10 GB. The 209 GB lives on disk and is read in the pieces that matter, which is what modern SSD bandwidth is good at.
What it costs
Speed. 1.59 tokens per second is roughly a slow typist. You would not hold a conversation at that rate. You would hand it something hard, go and do something else, and come back to an answer that a 9B model would not have produced.
That is the honest positioning. Plus is not the tier you chat with. It is the tier you escalate to.
What it needs
64 GB of unified memory is the real floor. Not because the weights need it, but because the paged engine needs the kernel to allow a large wired allocation, and the kernel caps that near half of physical memory. On a 32 GB Mac the cap is too low to hold what has to stay resident.
You also need 209 GB of free disk. On our own test machine that is currently the binding constraint rather than the memory, which is why our benchmark table lists Plus as unmeasured on MMLU while still reporting its decode speed. We would rather say that plainly than quietly leave the row blank.
Method
- Peak resident memory measured 2026-06-12 on an M1 Ultra Mac Studio, 64 GB, running the shipping V9 paged engine: 10 GB.
- Decode speed 1.59 tokens per second, same machine and engine. Speed varies by Mac; SSD bandwidth matters more here than for tiers that fit in memory.
- Model: Qwen3.5-397B-A17B, top-10 routing of 512 experts, MLX 4-bit. 209 GB on disk.
- Minimum memory in our catalog is 64 GB, set from the kernel wired-memory limit rather than from the model size.
All of this happens on your own machine with no network. Outlier ships smaller tiers for everyday work and Plus for the questions that deserve the wait. The FAQ covers the rest, and support is one person.