Michael Limberger

Hardware requirements

Before you install anything, check what your Mac can actually run and whether this 14B local stack will feel usable.

Unified memory

Traditional PCs keep CPU RAM and GPU VRAM separate. A big model has to fit in dedicated graphics memory, which is often 8, 12, 24, or 48GB with a hard ceiling.

Apple Silicon shares one unified memory pool between CPU and GPU. On a 32GB Mac, the GPU can use that whole pool. There is no separate VRAM wall and no copy between pools. That is why a mid-range MacBook Pro can run models that would need an expensive discrete GPU on a PC.

Metal is Apple's GPU framework, and Ollama uses it automatically. No CUDA install, no extra driver dance. That simplicity is a real advantage here.

The 14B memory budget

This tutorial uses a 14 billion parameter model at Q4_K_M quantization. Weights are compressed from 16-bit floating point down to roughly 4 bits per parameter, so the file is about 9GB.

File size is not the whole story. At runtime you also need memory for the KV cache, which stores conversation context and grows with the context window. Approximate totals:

At 4K context:   roughly 9.5 GB total
        At 8K context:   roughly 10.7 GB total
        At 16K context:  roughly 12.5 GB total
        At 32K context:  roughly 15-16 GB total

macOS itself needs about 4 to 5GB, plus whatever else is open. Chrome tabs are heavy; Safari is lighter but still adds up.

16GB (base M1/M2/M3/M4): tight but functional. After macOS takes its share you may have 10 to 12GB free. That is enough for the model at 4K to 8K context if you close the browser and other heavy apps first. If macOS starts swapping to disk, speed collapses from roughly 10 tokens per second toward waiting a long time per reply. Verdict: it works, barely. Close everything else first.

24GB (M2/M3/M4 Pro): the sweet spot. Comfortable headroom for 16K context with a few background apps. Recommended minimum for regular use.

32GB+ (Pro/Max): ideal. You can run Q5_K_M for more quality with long context, or stay on Q4_K_M with plenty of spare room. Great experience at 14B with no forced compromise.

48GB+ (Max/Ultra): overkill for 14B. Consider a larger model such as qwen2.5-coder:32b if you have that much RAM.

Check the machine

Apple menu, then About This Mac. Confirm three things:

1. Chip: Must say M1, M2, M3, or M4 (with or without Pro/Max/Ultra).
           If it says Intel, stop here. This guide is not for you.
        
        2. Memory: The number next to "Memory" or "Unified Memory". This is
           your total. 16GB minimum, 24GB recommended.
        
        3. macOS version: Must be Sonoma (14) or newer. If you're on Ventura
           (13) or older, update first.

For a live view, open Activity Monitor (Applications, Utilities), Memory tab. If Memory Pressure is already yellow or red before Ollama starts, close apps first.

+----------------------------------------------------------+
|  The single biggest factor in your experience is RAM.    |
|  Not the chip variant, not the clock speed. RAM.         |
|  16GB will work. 24GB will be comfortable. 32GB+ ideal.  |
+----------------------------------------------------------+

Disk space

The model file is about 9GB. Ollama stores models under ~/.ollama/models/ by default. OpenClaw and dependencies need a couple more gigabytes. Budget at least 15GB free.

df -h /

Check the Available column. If the specs look good, install Ollama next.