Grux AI Download

Can I raise how much memory a local LLM can use on a Mac?

Yes. macOS lets the GPU use only part of a Mac's unified memory, holding roughly a quarter to a third of it back by default, varying from Mac to Mac, and the owner can raise that limit with one setting, iogpu.wired_limit_mb, until the next restart. On the 16 GB Apple M4 Mac mini measured for this page, Metal reported 11.84 GiB, 74 percent of its memory, with the setting at 0, which means the default is in force. A local model that is Too big or Tight at the Mac picker's budget can rate higher at a larger limit, and the table below works out which, for every Apple memory size and three larger limits, with the same fit rule as the Mac picker. The price is what is left for macOS and your other apps, which the table shows beside each limit. The setting is described in an open Ollama pull request, not in Apple's documentation, so it is a power user's lever and this page recommends no value.

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

What the default limit is

Apple's Metal reports a recommended maximum working set size for each GPU: the memory a model may use without affecting its runtime performance, in Apple's words. The Mac picker rates models against 75 percent of unified memory. One Mac was measured for this page, and it agrees with that share. No other Mac was measured.

What was readValue
MacApple M4 Mac mini, 16 GB, macOS 26.3.1
Metal's recommended maximum working set size11.84 GiB, 74 percent of its memory
The iogpu.wired_limit_mb setting0, so the default is in force
Read on2026-11-19, with no model loaded

How the limit is read and raised

Ollama's documentation does not describe this yet. An open pull request, ollama/ollama 19527, proposes the text. It says:

sysctl iogpu.wired_limit_mb               # 0 means macOS's default is in force
sudo sysctl iogpu.wired_limit_mb=<MB>      # MB = GB x 1024, until the next restart
sudo sysctl iogpu.wired_limit_mb=0         # back to the default

What a larger limit rates

Choose your Mac's memory. Each row is the Mac picker's rule with a larger share of memory as the limit. The first row is the picker's own 75 percent budget; your Mac's default may be a little lower or higher. The last column lists the models whose rating changes from that row, and the middle column is the memory left for macOS and everything else you run. The pull request warns that a limit set too high can cause heavy swapping, so a row with little left is a risk, not a target. Nothing here recommends a row.

How much memory does your Mac have?

8 GB Mac: at the Mac picker's budget a model may use about 6 GB.

LimitLeft for macOS and appsWhat changes from the first row
75 percent, the Mac picker's budgetabout 6 GB2 GBSame ratings as the Mac picker.
80 percent6.4 GB1.6 GBLlama 3.2 3B: Good fit to Great fit; Qwen 3 VL 4B: Too big to Tight
85 percent6.8 GB1.2 GBLlama 3.2 3B: Good fit to Great fit; Qwen 3 VL 4B: Too big to Tight; Qwen 3.5 9B: Too big to Tight
90 percent7.2 GB0.8 GBLlama 3.2 3B: Good fit to Great fit; Qwen 3 VL 4B: Too big to Tight; Qwen 3.5 9B: Too big to Tight

16 GB Mac: at the Mac picker's budget a model may use about 12 GB.

LimitLeft for macOS and appsWhat changes from the first row
75 percent, the Mac picker's budgetabout 12 GB4 GBSame ratings as the Mac picker.
80 percent12.8 GB3.2 GBQwen 3 VL 4B: Good fit to Great fit; Qwen 3.5 9B: Good fit to Great fit; Gemma 4 12B: Tight to Good fit
85 percent13.6 GB2.4 GBQwen 3 VL 4B: Good fit to Great fit; Qwen 3.5 9B: Good fit to Great fit; Gemma 4 12B: Tight to Good fit
90 percent14.4 GB1.6 GBQwen 3 VL 4B: Good fit to Great fit; Qwen 3.5 9B: Good fit to Great fit; Gemma 4 12B: Tight to Good fit

18 GB Mac: at the Mac picker's budget a model may use about 13.5 GB.

LimitLeft for macOS and appsWhat changes from the first row
75 percent, the Mac picker's budgetabout 13.5 GB4.5 GBSame ratings as the Mac picker.
80 percent14.4 GB3.6 GBNo model changes rating.
85 percent15.3 GB2.7 GBGrux Local 9B: Good fit to Great fit
90 percent16.2 GB1.8 GBGrux Local 9B: Good fit to Great fit

24 GB Mac: at the Mac picker's budget a model may use about 18 GB.

LimitLeft for macOS and appsWhat changes from the first row
75 percent, the Mac picker's budgetabout 18 GB6 GBSame ratings as the Mac picker.
80 percent19.2 GB4.8 GBGemma 4 12B: Good fit to Great fit; Qwen 3.8 27B: Too big to Tight
85 percent20.4 GB3.6 GBGemma 4 12B: Good fit to Great fit; Qwen 3.8 27B: Too big to Tight; Gemma 4 26B: Too big to Tight
90 percent21.6 GB2.4 GBGemma 4 12B: Good fit to Great fit; Qwen 3.8 27B: Too big to Tight; Gemma 4 26B: Too big to Tight; GPT-OSS 20B: Too big to Tight

32 GB Mac: at the Mac picker's budget a model may use about 24 GB.

LimitLeft for macOS and appsWhat changes from the first row
75 percent, the Mac picker's budgetabout 24 GB8 GBSame ratings as the Mac picker.
80 percent25.6 GB6.4 GBGemma 4 26B: Tight to Good fit
85 percent27.2 GB4.8 GBGemma 4 26B: Tight to Good fit; GPT-OSS 20B: Tight to Good fit
90 percent28.8 GB3.2 GBGemma 4 26B: Tight to Good fit; GPT-OSS 20B: Tight to Good fit

36 GB Mac: at the Mac picker's budget a model may use about 27 GB.

LimitLeft for macOS and appsWhat changes from the first row
75 percent, the Mac picker's budgetabout 27 GB9 GBSame ratings as the Mac picker.
80 percent28.8 GB7.2 GBNo model changes rating.
85 percent30.6 GB5.4 GBMistral Small 3.2 24B: Tight to Good fit; Qwen 3 Coder 30B: Too big to Tight
90 percent32.4 GB3.6 GBMistral Small 3.2 24B: Tight to Good fit; Qwen 3 Coder 30B: Too big to Tight

48 GB Mac: at the Mac picker's budget a model may use about 36 GB.

LimitLeft for macOS and appsWhat changes from the first row
75 percent, the Mac picker's budgetabout 36 GB12 GBSame ratings as the Mac picker.
80 percent38.4 GB9.6 GBGemma 4 26B: Good fit to Great fit; Qwen 3 Coder 30B: Tight to Good fit; Qwen 3.5 35B: Too big to Tight
85 percent40.8 GB7.2 GBGemma 4 26B: Good fit to Great fit; GPT-OSS 20B: Good fit to Great fit; Qwen 3 Coder 30B: Tight to Good fit; Qwen 3.5 35B: Too big to Tight
90 percent43.2 GB4.8 GBGemma 4 26B: Good fit to Great fit; GPT-OSS 20B: Good fit to Great fit; Qwen 3 Coder 30B: Tight to Good fit; Qwen 3.5 35B: Too big to Tight

64 GB Mac: at the Mac picker's budget a model may use about 48 GB.

LimitLeft for macOS and appsWhat changes from the first row
75 percent, the Mac picker's budgetabout 48 GB16 GBSame ratings as the Mac picker.
80 percent51.2 GB12.8 GBNo model changes rating.
85 percent54.4 GB9.6 GBNo model changes rating.
90 percent57.6 GB6.4 GBQwen 3 Coder 30B: Good fit to Great fit

96 GB Mac: at the Mac picker's budget a model may use about 72 GB.

LimitLeft for macOS and appsWhat changes from the first row
75 percent, the Mac picker's budgetabout 72 GB24 GBSame ratings as the Mac picker.
80 percent76.8 GB19.2 GBNo model changes rating.
85 percent81.6 GB14.4 GBLlama 3.3 70B: Tight to Good fit
90 percent86.4 GB9.6 GBLlama 3.3 70B: Tight to Good fit

128 GB Mac: at the Mac picker's budget a model may use about 96 GB.

LimitLeft for macOS and appsWhat changes from the first row
75 percent, the Mac picker's budgetabout 96 GB32 GBSame ratings as the Mac picker.
80 percent102.4 GB25.6 GBNo model changes rating.
85 percent108.8 GB19.2 GBNo model changes rating.
90 percent115.2 GB12.8 GBNo model changes rating.

What this page does not cover

Where to check this

ClaimSource
Metal reports a recommended maximum working set size for the GPUApple, recommendedMaxWorkingSetSize
The setting, macOS 14 or later, the reset at reboot, 0 for the default, and the measured caseOllama pull request 19527
Another fit calculator limits memory to a percentage in its estimatesllmfit pull request 1089
The fit rule and every model's serving memoryGrux Cookbook.swift at the pinned commit

Questions

How much of a Mac's memory can a local LLM use?
By default all but roughly a quarter to a third of its unified memory, varying from Mac to Mac. Metal reported 11.84 GiB, 74 percent, on the 16 GB Mac measured for this page; it reports the exact figure for each Mac.
How do I raise the GPU memory limit on a Mac?
On macOS 14 or later, set iogpu.wired_limit_mb to a number of megabytes with sudo sysctl. According to an open Ollama pull request the setting is lost at the next restart, and 0 sets the default again. It also says to leave several GB for macOS and other apps, because setting it too high can cause heavy swapping. Unload the model with ollama stop, or restart Ollama, before loading it again.
Is it safe to raise the limit?
Not without care. The Ollama pull request says to leave several GB for macOS and other apps, because setting it too high can cause heavy swapping, and that the setting is lost at the next restart. That is why the table shows the memory left beside each limit, and why this page recommends none.
Does a larger limit make a model faster?
It can, when a model that was split between CPU and GPU then fits on the GPU. The Ollama pull request measured a 27B model that split between CPU and GPU running fully on the GPU after the change. This page measures no speed.
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.