Grux AI Download

Which local LLM can a MacBook Pro or Mac Studio run?

Memory decides it. A MacBook Pro or Mac Studio can serve a model whose running footprint is under about three quarters of its unified memory, so each step up in memory leaves more room for the same model. Apple sells the MacBook Pro with 16 GB up to 128 GB and the Mac Studio with 36 GB up to 512 GB, and the table below rates every model in Grux AI's catalogue at each size with the app's own fit rule, from Great fit to Too big. Install Ollama, pull the model your row names, and it answers on the Mac with nothing sent anywhere. A newer chip changes how fast a model runs, not whether it fits.

Download Grux AI 3.3.1, free, for macOS 14 or later on Apple silicon.

Grux AI 3.3.1 · sources read 2026-11-19 · published 2026-11-19

By memory

Unified memoryWhere Apple sells itLargest Great or Good fitTightToo big
16 GBMacBook ProGrux Local 9B (Good fit)Gemma 4 12B (Tight)8 models, from Qwen 3.8 27B (Too big)
24 GBMacBook ProGemma 4 12B (Good fit)none8 models, from Qwen 3.8 27B (Too big)
32 GBMacBook ProQwen 3.8 27B (Good fit)Gemma 4 26B (Tight), gpt-oss 20B (Tight), Mistral Small 3.2 24B (Tight)4 models, from Qwen 3 Coder 30B (Too big)
36 GBMacBook Pro (M5 Max), Mac Studio (M5 Max)gpt-oss 20B (Good fit)Mistral Small 3.2 24B (Tight)4 models, from Qwen 3 Coder 30B (Too big)
48 GBMacBook Pro (M5 Pro, M5 Max), Mac Studio (M5 Max)Mistral Small 3.2 24B (Good fit)Qwen 3 Coder 30B (Tight)3 models, from Qwen 3.5 35B (Too big)
64 GBMacBook Pro (M5 Pro, M5 Max), Mac Studio (M5 Max)Qwen 3.5 35B (Good fit)none2 models, from Llama 3.3 70B (Too big)
96 GBMac Studio (M5 Ultra)Qwen 3.5 35B (Great fit)Llama 3.3 70B (Tight)1 model, from gpt-oss 120B (Too big)
128 GBMacBook Pro (M5 Max), Mac Studio (M5 Max)Llama 3.3 70B (Good fit)gpt-oss 120B (Tight)none
256 GBMac Studio (M5 Ultra)gpt-oss 120B (Great fit)nonenone
512 GBMac Studio (M5 Ultra)gpt-oss 120B (Great fit)nonenone

The ratings use the which LLM can my Mac run rule: a model's memory while it is serving, against about 75 percent of unified memory, in the app's own bands. The catalogue is the app's, read at commit 4b8d1b6. A memory size not in this table, such as 18 GB on an older MacBook Pro, falls between the rows around it, and the Mac picker on that page rates every size it lists, up to 128 GB.

Start one

brew install ollama
ollama pull <the model id from your row>     # the Mac picker prints each pull line
ollama run <the model id from your row>

What the table does not cover

Where to check this

ClaimSource
MacBook Pro memory: 16GB, 24GB, 32GB, and with M5 Pro or M5 Max up to 128GBApple, MacBook Pro tech specs
Mac Studio memory: M5 Max 36GB, configurable to 48GB, 64GB or 128GB; M5 Ultra 96GB, configurable to 256GB or 512GBApple, Mac Studio tech specs
The fit rule and every model's serving memoryGrux Cookbook.swift at the pinned commit
Ollama runs a model with one command on a MacOllama README

Questions

How much memory does a Mac need to run a 70B model?
Llama 3.3 70B needs about 60 GB while running in the app's catalogue, so it is Too big on 64 GB, Tight on 96 GB and a Good fit on 128 GB. That is the same rule as the Mac picker.
Is a Mac Studio better than a MacBook Pro for a local LLM?
At the same memory they fit the same models. The Mac Studio goes to 512 GB and the MacBook Pro to 128 GB, so the Studio reaches models a laptop cannot hold. This page quotes no speed.
Does an older MacBook Pro follow the same table?
Yes, by memory. An older MacBook Pro with the same unified memory fits the same models and runs them more slowly on an older chip.
Can a 16 GB MacBook Pro run a local LLM?
Yes. The table shows the largest model that fits well at 16 GB and the one that is Tight; the Mac picker lists the rest.
Download Grux AI 3.3.1 Read the source

Free, MIT licensed. macOS 14 or later, Apple silicon. 29.1 MB, signed and notarized by Apple. No account, no server, no subscription.