How much unified memory do I need on a Mac?
Apple Silicon shares one memory pool between CPU, GPU and the system, which makes capacity cheap and bandwidth the thing that separates the tiers.
The short answer
Computed from this tool’s default settings — model size and the rest as most people start. Change them below for your own case.
A 8B model at Q4_K_M needs 6.89 GB, so 16 GB is the smallest configuration that holds it. Step up a chip tier for bandwidth rather than capacity if speed matters more than size.
Mac Unified Memory Calculator
What each memory configuration actually runs, and how fast.
A 8B model at Q4_K_M needs 6.89 GB, so 16 GB is the smallest configuration that holds it. Step up a chip tier for bandwidth rather than capacity if speed matters more than size.
| Machine | Usable | Bandwidth | Speed |
|---|---|---|---|
| Raspberry Pi 5 · 8 GB | 4.80 GB | 17 GB/s | Will not fit |
| Jetson Orin Nano Super · 8 GB | 5.60 GB | 102 GB/s | Will not fit |
| iPhone · A19 · 8 GB | 3.60 GB | 68 GB/s | Will not fit |
| iPhone · A18 Pro · 8 GB | 3.60 GB | 60 GB/s | Will not fit |
| iPhone · A19 Pro · 12 GB | 5.40 GB | 77 GB/s | Will not fit |
| Snapdragon 8 Elite · 12 GB | 5.40 GB | 77 GB/s | Will not fit |
| Raspberry Pi 5 · 16 GB | 9.60 GB | 17 GB/s | 3 tok/s |
| Snapdragon 8 Elite · 16 GB | 7.20 GB | 77 GB/s | 14 tok/s |
| MacBook Air M4 · 16 GB | 12.0 GB | 120 GB/s | 21 tok/s |
| MacBook Air M4 · 24 GB | 18.0 GB | 120 GB/s | 21 tok/s |
| MacBook Pro M4 Pro · 48 GB | 36.0 GB | 273 GB/s | 49 tok/s |
| MacBook Pro M4 Pro · 64 GB | 48.0 GB | 273 GB/s | 49 tok/s |
| MacBook Pro M4 Max · 128 GB | 96.0 GB | 546 GB/s | 97 tok/s |
What each setting changes
Every input moves the result for a reason. This is what each one does and where to find the value for your own machine.
| Setting | Default | What it changes |
|---|---|---|
| Model size | 8B | 6 options, from 3B to 405B. |
| Context length | 8192 tokens | Anywhere from 1,024 to 131,072 tokens. |
| Quantisation | Q4_K_M — 4.5 bits/weight | 5 options, from Q8_0 — 8.5 bits/weight to Q3_K_M — 3.5 bits/weight. |
Real answers across model size
The same calculation run at a range of settings, with everything else left at its default. These are computed by the tool itself, not written by hand.
| Model size | Smallest Mac that fits | Model needs | Quantisation | Expected speed |
|---|---|---|---|---|
| 3B | 8 GB | 3.54 GB | Q4_K_M | 8 tok/s |
| 8B | 16 GB | 6.89 GB | Q4_K_M | 3 tok/s |
| 14B | 16 GB | 10.7 GB | Q4_K_M | 12 tok/s |
| 32B | 48 GB | 21.8 GB | Q4_K_M | 12 tok/s |
| 70B | 64 GB | 44.5 GB | Q4_K_M | 6 tok/s |
A 3B model at Q4_K_M needs 3.54 GB, so 8 GB is the smallest configuration that holds it. Step up a chip tier for bandwidth rather than capacity if speed matters more than size.
How this is calculated
No lookup tables and no invented constants. Here is the arithmetic, so you can check it against your own numbers.
Apple Silicon shares one memory pool between CPU, GPU and the operating system, so the usable fraction is lower than on a discrete card — macOS caps what a single GPU process may allocate, and the desktop needs its share.
Bandwidth, not capacity, is what separates the chip tiers. A Max part has roughly three to four times the bandwidth of a base part at the same capacity, and that ratio is close to the ratio of tokens per second you see.
These are well-founded engineering estimates, not benchmark results. Your quantisation, runtime and context length all move the real number, and usable memory is an assumption rather than a specification. See the full methodology for every assumption behind these figures.
Who this is for
The three situations that bring people to this calculation.
Configuring a Mac
Choose a memory tier from the models you want to run.
Comparing chips
See what a Pro or Max buys in speed.
Existing Mac
Find what your machine already handles.
How to use this calculator
Four steps, no account, nothing leaves your browser.
Set model size
Start at the top of the panel. Every figure recalculates as you change it — there is no submit button, because watching the number move is the point.
Adjust the rest to match your setup
2 further settings: context length, quantisation. Defaults are the common case, so change only what differs for you.
Read the headline, then the table
The large figure answers the question. The table underneath shows how the answer changes across nearby settings, which is usually where the decision actually gets made.
Check it against the method
The arithmetic is written out above. If a number looks wrong for your hardware, the assumptions are the first place to look — usable memory and quantisation are the two that vary most.
How much unified memory do I need on a Mac: common questions
The questions people ask about this, answered without hedging.
Is 16GB enough to run local LLMs on a Mac?
For 7–8B models at 4-bit, yes — about 5 GB of weights plus cache leaves room to work in. For 32B you want 32GB or more, and 70B needs 64GB before it is comfortable.
How much unified memory do I need for local AI?
16GB runs 7–8B models comfortably at 4-bit. 32GB opens 32B. 64GB and above is where 70B becomes practical. Memory is not upgradeable later, so buy ahead of what you need.
Why can I not use all my unified memory?
macOS reserves a share for the system and caps what a single GPU process may allocate. The usable fraction is meaningfully lower than the number on the spec sheet.
Is a Mac good for running LLMs?
Very good on capacity per dollar and excellent on power draw. A Max chip has bandwidth competitive with mid-range discrete cards; a base chip is much slower.
Base, Pro or Max?
For capacity alone, memory size is what matters. For speed, the chip tier is what matters — a Max has roughly three to four times a base chip’s bandwidth, and tokens per second follow that ratio closely.
Related local AI tools
All 50 run on the same arithmetic, so answers across them agree.
More on hardware
Will it fit
Context
Speed
Quantisation
Runtime & errors
Cost
Searches this page answers
Different ways of asking the same question, all resolved above.
Now find the models that fit
Sizing is only half the problem. Model Radar takes your hardware and shows which models actually run on it, ranked by what they are good at — the same arithmetic as this page, applied to every model worth running.