I Moved My AI Stack From NVIDIA to Apple Silicon. Then I Moved Back. Here's Why.

I bought a Mac Mini M4 thinking it would replace my Windows PC for local AI. Spoiler: it didn't. But it also didn't fail — it just turned out to be good at different things than I expected.

Here's six months of actual usage data, with the real numbers nobody puts in benchmark charts.

The Hardware

Machine	CPU/GPU	RAM	What I Paid
Mac Mini M4	10-core M4, 10-core GPU	16GB unified	$599
Windows PC	AMD 9970X	128GB	~$2,500 (existing)
	RTX 3060 12GB		$150 (used)

The Mac Mini is silent, tiny, and sips power. The PC sounds like a jet engine under load and pulls 200W+ when the GPU is working. I wanted the Mac to replace the PC for everything AI-related. I was half right.

The Migration (Month 1-2)

I installed Ollama on the Mac Mini, pulled Qwen 3.5 9B, and started using it as my daily driver. The first thing I noticed: it's fast. Really fast. For a 9B model, inference felt snappier than the RTX 3060 running the same model.

But then I tried loading bigger models.

Qwen 3 Coder 30B (Q4_K_M, ~18GB): Loaded fine on the RTX 3060 (12GB VRAM... wait, that shouldn't work). Oh right — it offloads to system RAM. Slow, but functional. On the Mac Mini? Out of memory. 16GB unified memory isn't 16GB free memory — the OS, browser, and apps eat 6-8GB before Ollama even starts.

DeepSeek R1 8B: Runs great on both. Mac Mini: ~25 tok/s. RTX 3060: ~35 tok/s. The GPU wins, but not by enough to matter for interactive use.

Qwen 3.5 9B: Mac Mini: ~18 tok/s. RTX 3060: ~28 tok/s. Again, GPU is faster, but both feel instant.

So far: Mac Mini handles small models well. But I hit the wall fast.

The Wall (Month 3)

Three things broke my "Mac only" dream:

1. Memory Pressure

16GB unified memory sounds like a lot until you're running:

• Ollama (4-8GB for a loaded model)
• VS Code with a few extensions (2-3GB)
• Safari with 20 tabs (3-4GB)
• A terminal, a music player, maybe a video call (2GB)

That's 11-17GB before you even ask the model a question. macOS starts swapping. Ollama slows down. The whole machine gets sluggish.

On the Windows PC with 128GB RAM? I can load a 30B model and run a game and compile code simultaneously. The RTX 3060's 12GB VRAM is a separate pool that doesn't compete with the OS.

2. Model Availability

Not every model runs well on Apple Silicon. The M4 has excellent support for mainstream models (Llama, Qwen, Mistral), but niche or newly released models often ship with "CUDA only" initial releases. I spent an afternoon trying to get a specific vision model running on the Mac before giving up and running it on the PC in 10 minutes.

The NVIDIA ecosystem is still the default for AI tooling. Apple Silicon is catching up, but it's not there yet.

3. Multi-User / Multi-Model

I wanted to run two models simultaneously — a coding assistant and a general chatbot. On the Mac Mini, loading two 9B models (~12GB total) leaves zero headroom for the OS. Everything crawls.

On the PC, I can load a 30B model in VRAM and a 7B model in system RAM, and both respond quickly. The GPU handles the heavy one; the CPU handles the light one. Separate memory pools are a feature, not a bug.

What the Mac Mini Actually Excels At

I didn't move everything back to the PC. The Mac Mini found its niche, and it's a good one:

Always-On Tasks

The Mac Mini draws ~15W at idle. The PC draws ~80W at idle (that GPU doesn't sleep well). For 24/7 tasks — routing, scheduling, lightweight notifications — the Mac wins on power cost alone.

Small Model Inference

For anything 9B and under, the Mac Mini is genuinely competitive. The unified memory architecture means there's no "copy to VRAM" overhead. A 4B model on the Mac can feel faster than the same model on the GPU because the latency is lower.

No Driver Drama

NVIDIA drivers on Windows are... fine. Until they aren't. An update breaks CUDA, a new Ollama version wants a different driver, and suddenly you're debugging nvidia-smi on a Tuesday night. The Mac just works. I have never, in six months, had an Apple Silicon AI setup break because of a system update.

Portability

I unplugged the Mac Mini, took it to a different room, plugged it back in, and my entire AI stack was back online in 30 seconds. Try that with a full tower PC.

The Real Numbers

Here's a month of actual usage, tracked with a simple script that logs tokens/sec and power draw:

Task	Mac Mini M4	RTX 3060	Winner
Qwen 3.5 9B, 500-token prompt	18 tok/s	28 tok/s	RTX 3060
DeepSeek R1 8B, reasoning	25 tok/s	35 tok/s	RTX 3060
Qwen 3 4B, quick chat	42 tok/s	55 tok/s	RTX 3060
30B model loading	❌ OOM	✅ Loads (slow)	RTX 3060
Two models simultaneously	❌ Crawls	✅ Works	RTX 3060
Idle power draw	~15W	~80W	Mac Mini
Load power draw	~35W	~220W	Mac Mini
Setup time (new model)	2 min	2 min	Tie
Driver reliability	Flawless	Occasional issues	Mac Mini

What I Actually Do Now

I stopped trying to pick a winner. Both machines have jobs:

Mac Mini M4:

• Always-on orchestration (routing, scheduling, notifications)
• Small model inference (4-9B) for quick tasks
• Coding and development (it's my daily driver)
• Runs 24/7, costs ~$3/month in power

Windows PC + RTX 3060:

• Heavy inference (30B models, image generation)
• GPU rental on Vast.ai when I'm not using it
• Anything that needs CUDA or more than 8GB VRAM
• On-demand, not 24/7

This isn't a benchmark conclusion. It's a workflow conclusion. The Mac Mini is the brain. The PC is the muscle. Separating them by task type eliminated the "which machine should I use?" decision fatigue and made both faster at what they're good at.

The Honest Bottom Line

If you're buying one machine for local AI and you want to run large models: get an NVIDIA GPU. Full stop. The ecosystem, the memory architecture, and the raw inference speed are still unbeatable for serious workloads.

If you want a silent, efficient, always-on machine for smaller models and orchestration: the Mac Mini M4 is excellent. Just don't expect it to replace a dedicated GPU for everything.

And if you already have both? Stop trying to consolidate. Use each for what it's good at. The separation is the feature.

---

Sam Hartley is a solo dev running a split AI stack across a Mac Mini and a Windows PC. No benchmarks were harmed in the making of this article — just a lot of time commands and a power meter.

Drop your setup in the comments — curious if others have found the same split, or if you've made a single-machine setup work for everything.