How much RAM does a Mac need for local LLMs?

16 GB runs 7–9B models, 24–36 GB runs 14B and some 27B models, 48–64 GB runs 32B models comfortably and 70B at Q4 with short context, and 128 GB runs 120B-class MoE models. Buy the most memory you can; it cannot be upgraded later.

Is M4 Pro or M4 Max better for LLMs?

The Max, if you can afford it: it has twice the bandwidth (546 vs 273 GB/s), so it generates about twice as fast, and it goes up to 128 GB.

Can a Mac run a 70B model?

Yes with 64 GB or more. Llama 3.3 70B needs about 43.2 GB at Q4_K_M, which fits the roughly 48 GB a 64 GB Mac gives the GPU only with a short context, and comfortably on 96–128 GB Macs.

How much unified memory can the GPU use on a Mac?

By default about two thirds of memory on 16–36 GB Macs and about three quarters on larger ones. You can raise it with sudo sysctl iogpu.wired_limit_mb until the next reboot.

Is the M3 Ultra or M4 Max better for LLMs?

The M3 Ultra has much more memory (up to 512 GB) and about 50% more bandwidth, so it runs larger models and runs them faster. The M4 Max is far cheaper and enough for models up to about 90 GB.

Best Mac for Local LLMs

Updated October 14, 2026 · Figures calculated live from our model data

On a Mac, unified memory is your VRAM, but macOS only lets the GPU use part of it: about two thirds on 16–36 GB machines and about three quarters on larger ones. A 128 GB Mac has about 96 GB for models. Bandwidth then decides speed: a Max chip is roughly twice as fast as a Pro with the same memory.

What fits on each Mac

Largest notable models that fit fully at Q4_K_M (or native format) with an 8K context, using the default GPU memory limit.

Mac	GPU-usable	Bandwidth	Runs well	8B speed
Apple M4 16GB	10.7 GB	120 GB/s	Qwen 3.5 9B (6.4 GB)	~14 tok/s
Apple M4 Pro 24GB	16.0 GB	273 GB/s	Qwen 3 14B (10.5 GB)	~32 tok/s
Apple M4 Max 36GB	24.0 GB	410 GB/s	Qwen 3.6 27B (16.9 GB)	~49 tok/s
Apple M4 Pro 48GB	36.0 GB	273 GB/s	Qwen 3 32B (21.4 GB)	~32 tok/s
Apple M5 Pro 64GB	48.0 GB	307 GB/s	Llama 3.3 70B (43.2 GB)	~36 tok/s
Apple M4 Max 128GB	96.0 GB	546 GB/s	gpt-oss-120b (62.0 GB)	~64 tok/s
Apple M5 Max 128GB	96.0 GB	614 GB/s	Qwen 3.5 122B-A10B (70.2 GB)	~72 tok/s
Apple M2 Ultra 192GB	144 GB	800 GB/s	Qwen 3 235B-A22B (135 GB)	~93 tok/s
Apple M3 Ultra 512GB	384 GB	819 GB/s	Qwen 3.5 397B-A17B (226 GB)	~95 tok/s

Raising the GPU memory limit

You can let the GPU use more memory until the next reboot:

# example for a 128 GB Mac: allow 112 GB for the GPU
sudo sysctl iogpu.wired_limit_mb=114688

Leave at least 8 GB for macOS and your apps, more if you run other heavy software. This turns a "tight" fit into a comfortable one, for example Llama 3.3 70B at Q8_0 (73.2 GB) on a 128 GB machine.

Mixture-of-experts models are the Mac's sweet spot

Macs have lots of memory but less bandwidth than high-end GPUs. MoE models only read their active parameters per token, so they suit Macs well: gpt-oss-120b generates ~83 tokens/s on an Apple M4 Max 128GB, while the dense Llama 3.3 70B manages ~8.0.

Mac vs NVIDIA

Capacity: a 128 GB Mac holds models that need three or four 24 GB GPUs.
Generation speed: an RTX 5090 32GB is about three times faster on models that fit in 32 GB.
Prompt processing: much slower on Macs before the M5 generation, which matters for long documents and coding agents. M5 chips add neural accelerators to each GPU core that narrow the gap.
Image generation and training: far slower than NVIDIA; most tools are built for CUDA first.
Noise and power: a Mac Studio running a 120B model draws a fraction of a multi-GPU PC.

Compare directly: M4 Max vs RTX 5090, M4 Max vs DGX Spark, M4 Max vs M5 Max.

Which runtime?

LM Studio (GGUF and MLX), Ollama and llama.cpp all use the GPU through Metal. MLX builds are often the fastest for generation on Apple Silicon; GGUF has the widest model selection.

Mac mini, MacBook Pro or Mac Studio?

Mac mini: the cheapest way in. The M4 Pro configurations with 48–64 GB run 32B models well, but bandwidth (273 GB/s) limits speed.
MacBook Pro: the only portable option with Max chips and up to 128 GB. Fans get loud under sustained load.
Mac Studio: Max or Ultra chips, the most memory and the best cooling. The 512 GB Apple M3 Ultra 512GB is the largest single-box option for local AI.

Example speeds on one Mac

Estimated generation speed on an Apple M4 Max 128GB, at Q4_K_M or native format with 8K context:

Model	VRAM	Speed
Llama 3.1 8B	6.3 GB	~64 tok/s
Qwen 3 14B	10.5 GB	~36 tok/s
Qwen 3.6 27B	16.9 GB	~21 tok/s
Qwen 3.6 35B-A3B	21.0 GB	~105 tok/s
Llama 3.3 70B	43.2 GB	~8.0 tok/s
gpt-oss-120b	62.0 GB	~83 tok/s
Qwen 3.5 122B-A10B	70.2 GB	~44 tok/s

Prompt processing: the hidden cost

Generation speed is only half the story. Before a model writes anything, it reads your whole prompt. On NVIDIA GPUs this is very fast; on Macs before M5 it is several times slower. For a short chat message it does not matter. For a 30K-token coding-agent prompt it can mean waiting a minute before the first word. M5 chips add neural accelerators to each GPU core, which speeds this step up considerably.

Tips for running models on a Mac

Close memory-heavy apps; macOS will compress or swap memory, which slows everything down.
Keep at least 8 GB free for the system if you raise the GPU memory limit.
Use MLX builds in LM Studio where available.
Keep the Mac plugged in; laptops may throttle on battery.

Is a Mac good for image generation?

It works with ComfyUI and Draw Things, but it is several times slower than a mid-range NVIDIA card for FLUX and SDXL. If images or video are your main use, a GPU with 16–24 GB is the better buy.