Free tool
How much unified memory do you need?
A 16 GB Mac does not give a model 16 GB. Unified memory is shared with the operating system, and the GPU only gets a share of it, which is why the spec on the box is never the number that decides whether a model loads. Pick what you want to run and see the honest minimum.
Llama 3.1 8B at Q4_K_M, 8K context. Needs about 12.2 GB. Cheapest machine that fits: 16 GB.
The workhorse. Near-lossless for chat, runs on the least memory.
Buy at least
16 GBunified memory
The model itself needs 6.2 GB, but unified memory is shared: the GPU is typically capped near 70% of the total and the OS still needs its own room. That is why the honest requirement is 12.2 GB, not 6.2.
- Weights
- 4.5 GB
- KV cache
- 1 GB
- Overhead
- 0.7 GB
- 8 GBWon't load
- 16 GBCheapest that fits
- 18 GBRoom to spare
- 24 GBRoom to spare
- 32 GBRoom to spare
- 36 GBRoom to spare
- 48 GBMore than this needs
- 64 GBMore than this needs
- 96 GBMore than this needs
- 128 GBMore than this needs
- 256 GBMore than this needs
- 512 GBMore than this needs
Estimates, not benchmarks. Real usage moves with the runtime, the exact model architecture, and how much else you have open. On a Mac you can raise the GPU’s share above the default with iogpu.wired_limit_mb, which buys headroom on a tight machine, at the risk of the OS swapping.
Know the number. Now pick the machine that has it.
Unified memory vs VRAM, briefly
On a machine with a discrete graphics card, a model has to fit inside that card’s VRAM. Exceed it and the model spills into system memory, where generation slows to a crawl. The ceiling is hard, and on consumer cards it is low: 16 GB on an RTX 5080, 32 GB on a 5090.
Unified memory removes that wall. The CPU and GPU address one pool, so a 128 GB machine can hold a 70B model that no consumer graphics card can. What you trade is bandwidth: dedicated GDDR7 and HBM move data faster than unified memory does, so when a model fits a real GPU, the GPU usually wins on tokens per second. The rule that follows is simple: unified memory buys you capability, a discrete GPU buys you speed.
If you are choosing hardware rather than sizing it, the GPU checker covers the discrete side, and the laptop guide names the actual machines.
Frequently asked
How much unified memory do I need for local AI?
For an 8B model at Q4 with a normal 8K context, 16 GB is the realistic floor. The reason the number is higher than the model size is that unified memory is shared: the GPU is typically capped near 70% of the total, and macOS plus a browser and an editor still need their own room. As a rule of thumb, take the model's memory footprint and roughly divide by 0.7, or add about 6 GB, whichever is larger. For a 70B model at Q4 you are looking at 96 GB and up.
Do I need 16GB or 24GB of unified memory?
Take 24 GB if you intend to run local models at all seriously. 16 GB will load a 7-8B model at Q4 and works fine for chat, but it leaves almost nothing for a long context or for the apps you actually work in, so you will be closing things to make room. 24 GB clears that ceiling: an 8B model with a 32K context asks for 16 GB and a 14B at Q4 asks for 18 GB. If local AI is the reason you are buying the machine, the jump from 16 to 24 GB is the single highest-value upgrade on the configuration page.
Which is better, RAM or unified memory?
They are not competing things: unified memory IS the RAM, shared between the CPU and GPU instead of split into separate pools. The advantage for AI is real: on a discrete-GPU machine a model must fit in VRAM (often 8-16 GB) and anything larger spills to system RAM and collapses in speed, while on a unified-memory machine the GPU can address most of the whole pool, so a 128 GB Mac can run models that no consumer graphics card can hold. The trade-off is bandwidth: unified memory is slower than dedicated GDDR or HBM, so a model that fits a real GPU will usually generate tokens faster there.
Why can't a model use all of my unified memory?
Because the operating system reserves a share and the GPU is given a wired limit rather than the whole pool. On Apple Silicon the default allows roughly 67-75% of unified memory to be claimed by the GPU, which is why a 16 GB machine realistically offers a model around 11.2 GB. You can raise the ceiling on macOS with the iogpu.wired_limit_mb setting, which buys headroom on a tight machine, but pushing it too far starves the OS and the system starts swapping, which is slower than the problem you were solving.
How much unified memory do I need to run a 70B model?
96 GB at a normal context, and 128 GB if you want the full 128K. A 70B model at Q4_K_M is 39.2 GB of weights alone, plus a KV cache that grows quickly with context and the runtime's own overhead. Divided by the GPU's share of unified memory, that comes to about 68 GB, which is past a 48 GB machine entirely rather than merely tight on one. This is the range where unified memory genuinely beats consumer graphics cards, since no single consumer GPU carries enough memory for a 70B model at all.
Does context length change how much memory I need?
Yes, and it is the part people forget. The weights are a fixed cost, but the KV cache grows linearly with the number of tokens in the context and scales with the model's layer count, so moving a 70B model from an 8K to a 128K context moves the machine you need from 96 GB to 128 GB. If a machine is borderline for the model you want, cutting the context length is usually the cheapest way to make it fit.