Orbital Desk.

AI Tools

Can My Mac Run This Model?

Running a large language model locally is a memory question before it is anything else. Pick your Mac, pick a model, and get a straight answer: runs comfortably, runs tight, or needs a cluster of Macs pooling their memory.

…
verdict
…
memory the model needs
…
usable on your Mac
…
headroom

How the math works

A model's weights take roughly its parameter count times the bytes per parameter: about 0.56 GB per billion parameters at 4-bit, 1.05 GB at 8-bit, and 2 GB at full precision. A 70B model at 4-bit is therefore about 39 GB before it answers a single token.

On top of the weights sit the KV cache (the model's working notes on your conversation, which grows with context length and is counted here at full precision) and about 1.5 GB of runtime. Against that, a Mac cannot give a model all of its unified memory: macOS and the GPU memory ceiling leave roughly 75% of RAM usable for one model. If the total fits inside that 75%, it runs. If it fits inside total RAM but not the usable share, expect memory pressure and slowdowns.

Too big for one machine? New clustering tools pool several Macs over Thunderbolt 5 and treat their combined memory as one pool, minus a system reserve on each machine. The cluster count above uses that math. It is how a model like the full DeepSeek R1, about 376 GB at 4-bit, becomes a four-Mac problem instead of an impossible one.