MacforAIHow Much RAM Do You Need for a Local LLM on a Mac?
Watch the video on YouTube
A Mac has no separate video memory. The CPU and the GPU share one pool of unified memory, and a local LLM has to fit in it together with macOS and your other apps. That is why RAM, not the chip name, is the first thing to check when you pick a Mac for local AI.
What actually takes memory
Three things add up when a model is loaded:
- Weights. Roughly
parameters × bits per weight ÷ 8. An 8B model at 4-bit is about 4.5–5 GB. A 70B model at 4-bit is about 40 GB. At 8-bit the numbers double. - KV cache. The model keeps keys and values for every token of the conversation so it does not recompute them. It grows linearly with context length: a short chat costs almost nothing, a 128K-token context can add several gigabytes on a larger model.
- Buffers and runtime overhead. Compute buffers, the app itself (Ollama, LM Studio, llama.cpp, MLX) and macOS. Plan for a gigabyte or two.
Not all of your RAM is available to the GPU
macOS limits how much unified memory the GPU may wire down. As a rule of thumb it is a bit over two thirds of RAM on smaller configurations and about three quarters on larger ones. A 16 GB Mac therefore gives a model roughly 10–11 GB, not 16. Advanced users can raise the limit, but the system gets less headroom and can become unstable, so treat that as a last resort.
Rough RAM targets
These are ballpark figures for 4-bit quantized models and a moderate context. Use them as a starting point, not a guarantee.
| Unified memory | Usable by the GPU | Comfortable model size |
|---|---|---|
| 16 GB | ~10–11 GB | up to ~8B |
| 24 GB | ~16 GB | up to ~14B |
| 32 GB | ~22 GB | ~14B with long context, or ~30B-class models |
| 64 GB | ~48 GB | 70B-class at 4-bit |
| 128 GB and up | ~96 GB and up | large mixture-of-experts models, 70B at 8-bit |
Mixture-of-experts (MoE) models are the exception to "size equals speed": all experts must sit in memory, so they need as much RAM as their total size, but each token only reads the active experts, so they generate faster than a dense model of the same size.
Quantization and context are your levers
If a model is just over the limit, you have two cheap options: choose a lower-bit quantization (4-bit instead of 8-bit halves the weights with a modest quality loss), or cut the context window. Both are visible in the calculator, which shows the exact split between weights, KV cache and buffers for the model and Mac you choose.
The short version
- Pick the model first, then the RAM. Memory is the one thing you cannot upgrade later.
- Multiply by 0.65–0.75 to see what the GPU can really use.
- Leave room for context. Long documents and agents need it.