The question I get most: "What computer should I buy for local AI?" After running models on multiple machines, here's the definitive guide by budget tier.
The One Rule
Memory is everything. Not the processor, not the clock speed. When model weights exceed available memory, performance drops 97% — from usable to unusable. Every recommendation below is built around maximizing AI-accessible memory per dollar.
Budget Tier: Under $600
Mac Mini M4 16GB — $499
The entry point. Runs 7B models at 30+ tokens per second. Enough for a useful local assistant and coding helper. Limitation: 16GB fills up fast if you run anything alongside the model.
Beelink SER8 / GEEKOM A9 — 32GB DDR5 (~$500-550)
AMD Ryzen mini PCs running Linux. 7B-13B models at 20-25 tokens per second. Cheaper than Mac, RAM is upgradeable. Trade-off: Linux + ROCm setup is more complex than macOS + Ollama.
Sweet Spot Tier: $800-1,500
Mac Mini M4 32GB — $799 ← My Pick
This is what I run. 32GB unified memory handles 7B-13B models with room for ComfyUI, n8n, and OpenClaw running simultaneously. Replaces $80-100/month in cloud subscriptions, pays for itself in under a year. Silent, tiny, 15 watts idle. Full setup in my Mac Mini AI Server guide.
Used RTX 3090 Build — ~$1,200-1,500
24GB VRAM, used cards at $650-750. Pair with a basic system for a total around $1,300. Runs 13B-30B models with full CUDA acceleration. 3-5x faster than the Mac for image generation. Trade-off: louder, hotter, 350W under load.
Choose over the Mac if image generation speed is critical or you need CUDA-specific workloads.
Power User Tier: $2,000-3,000
Mac Mini M4 Pro 64GB — $2,199
Runs 30B-40B models at 12-18 tokens per second. These models are dramatically more capable than 7B-8B for complex reasoning. Memory bandwidth of 273 GB/s translates directly to faster inference. Worth it if you work with larger models daily.
Minisforum MS-S1 Max 128GB — ~$2,300-3,000
The monster. AMD Ryzen AI Max+ with 128GB unified memory. Can hold a full 70B model in RAM. The most capable local AI machine outside a multi-GPU workstation. Trade-off: ROCm is less mature than Apple's Metal, more setup friction.
RTX 4090 Build — ~$2,500-3,000
Fastest single consumer GPU for AI inference. 50+ tokens per second on quantized models, unmatched image generation speed. 24GB VRAM limits full 70B models but dominates everything under 30B.
What NOT to Buy
Any GPU under 12GB VRAM. 8GB cards fill up immediately. Useless for AI despite being fine for gaming.
A Mac with 8GB RAM. Exists, is cheap, can't run anything useful.
A cloud GPU subscription for daily inference. Local hardware is cheaper within 3-8 months and you own it forever.
My Recommendation by Use Case
What to Buy
Just trying local AI: Mac Mini M4 16GB ($499)
Building an AI stack: Mac Mini M4 32GB ($799) — what I run
Fast image generation: Used RTX 3090 build (~$1,300)
Running 30B+ models: Mac Mini M4 Pro 64GB ($2,199)
Maximum capability: Minisforum MS-S1 Max 128GB (~$2,500)
Don't overthink it. Buy the most memory you can afford. Install Ollama. Pull a model. Start building. The skills you develop transfer to any hardware you upgrade to later.