THE RIG — COMPARISON
Best Local LLM Hardware
Mini PCs vs. GPUs for running Ollama and local AI agents. Benchmarked on llama3.2 3B Q4_K_M.
Agent Score (0–100):
A practical rating for local-agent inference. 80–100 means desktop-class VRAM and fast CUDA/ROCm throughput. 40–60 means usable for light agents but not for large models or long contexts.
| Product | Type | Price | Agent Score | Best For | Action |
|---|---|---|---|---|---|
| Geekom A8 | mini-pc | $899 | 52 tok/s | Entry to mid-range local inference | View → |
| Beelink SER7 Pro | mini-pc | $649 | 42 tok/s | Entry to mid-range local inference | View → |
| Intel NUC 13 Pro | mini-pc | $$480 | tok/s | Entry to mid-range local inference | View → |
| NVIDIA RTX 4090 NVIDIA | gpu | $1599 | 95 tok/s | Large models / maximum speed | View → |
| AMD Radeon RX 7900 XTX | gpu | $999 | 78 tok/s | Large models / maximum speed | View → |
| AMD RX 7900 XT | gpu | $$899 | tok/s | Large models / maximum speed | View → |
| NVIDIA RTX 4080 SUPER | gpu | $$999 | tok/s | Large models / maximum speed | View → |
FAQ
What is the best mini PC for running local LLMs?
The Beelink SER7 Pro is the best value for most users, offering ~42 tok/s on llama3.2 3B. For more headroom, the Geekom A8 with Ryzen 9 8945HS reaches ~52 tok/s.
Do I need a dedicated GPU for local AI inference?
No. Modern AMD Ryzen mini PCs with RDNA 3 integrated graphics can run models up to 7B-13B parameters. A dedicated GPU like the RTX 4090 is only needed for large models or maximum speed.
How much RAM do I need for Ollama?
32GB is the sweet spot for 7B models. 64GB lets you run 13B models comfortably or multiple smaller models at once.