Quick pick: For most people who want a small, quiet box running Ollama 24/7, a Ryzen mini PC with upgradeable DDR5 and 32GB to 64GB of RAM (like the Beelink SER8) is the value play. If you want to run 30B to 70B models locally, you need wide unified memory: the GMKtec EVO-X2 (Ryzen AI Max+ 395, up to 128GB) or a Mac mini with a lot of memory.
Every spec below comes from the manufacturer's own product page. We do not list prices because they change daily; use the links to check the current one.
What actually matters in a local AI mini PC
- Memory capacity decides which models fit. See our RAM and VRAM chart for Ollama.
- Memory bandwidth decides how fast replies come out. A normal dual channel DDR5 mini PC is fine for 8B models; unified memory designs are much faster on big ones.
- Upgradeable vs soldered RAM. Socketed RAM lets you start at 32GB and grow. Soldered means buy the size you will need on day one.
- Noise and power, if it sits on your desk or runs all day as a home server.
The picks
Beelink SER8 (Ryzen 7 8845HS): best value, upgradeable
- CPU: AMD Ryzen 7 8845HS, 8 cores / 16 threads, up to 5.1GHz
- GPU: Radeon 780M integrated (12 cores)
- RAM: DDR5 5600 in SO-DIMM slots, Beelink lists support up to 256GB
- Storage: dual M.2 2280 PCIe 4.0 slots
- Ports: USB4, HDMI 2.1, DisplayPort 1.4, 2.5G LAN
- 65W TDP, Beelink rates it around 32dB
Why it works: Socketed RAM is rare in this size. Start with 32GB, move to 64GB later, and run 8B to 14B models all day at low power. Great as a home Ollama and Open WebUI server.
Downsides: Dual channel DDR5 is the bottleneck. 30B dense models will run but slowly, and 70B is not practical. The integrated GPU helps little with Ollama on Windows, so think of this as a CPU inference box.
Skip it if: you want 30B+ models at chat speed, or you want to game seriously on the same machine.
GMKtec EVO-X2 (Ryzen AI Max+ 395): for big models
- CPU: AMD Ryzen AI Max+ 395, 16 cores / 32 threads, up to 5.1GHz
- GPU: Radeon 8060S integrated, 40 RDNA 3.5 compute units
- RAM: 64GB or 128GB LPDDR5X 8000 (soldered, not upgradeable); GMKtec also lists a 96GB variant
- Storage: dual M.2 2280 PCIe 4.0 slots
- Ports: USB4 x2, HDMI 2.1, DisplayPort 1.4, 2.5G LAN, Wi-Fi 7
- Runs at 120W sustained, 140W peak, about 230W power brick
Why it works: This is the class of mini PC that can hold 70B and even 100B+ MoE models in memory. GMKtec's own benchmark lists about 57 tokens/s on gpt-oss:20b and about 19 tokens/s on gpt-oss:120b on the 128GB model (manufacturer figures, your results will vary).
Downsides: Expensive, RAM cannot be upgraded, and owner reviews repeatedly mention loud fans under CPU load. You may need a BIOS update to raise the GPU memory allocation.
Skip it if: you only plan to run 8B models. The SER8 or your current PC with more RAM does that for far less money.
Apple Mac mini (M6 or M5 Pro): quietest, easiest
- Current lineup per Apple: M6 with 16GB, 24GB or 32GB unified memory (153GB/s to 170GB/s bandwidth)
- M5 Pro with 24GB, 48GB or 64GB unified memory (307GB/s)
- 5 x 5 x 2 inch body, Apple lists 5 dBA at idle
- Ollama has a native macOS app
Why it works: Unified memory acts like VRAM, so a 32GB M6 handles 14B to 20B models smoothly and a 64GB M5 Pro handles 30B to 70B 4-bit models. Near silent and very low power.
Downsides: Memory is fixed at purchase and Apple charges a lot for upgrades. Base 16GB models are only good for 8B class models. No CUDA, so some Windows and NVIDIA focused tools will not run.
Skip it if: you need Windows, or you want the most memory per dollar.
Which one should you get?
- Private chat, writing help, 8B to 14B models, home server: SER8 class mini PC with 32GB to 64GB RAM.
- Mac user, wants silence: Mac mini with at least 24GB, ideally 32GB.
- Wants 70B models or big MoE models locally: EVO-X2 128GB or Mac mini M5 Pro 64GB.
- Already own a desktop tower: do not buy a mini PC. Add RAM or a GPU instead; see best budget GPU for Ollama and budget AI PC under $1,000.
Once it arrives, follow our Ollama + Open WebUI install guide.
FAQ
Can a mini PC run local AI?
Yes. Any modern mini PC with 16GB of RAM can run 7B to 8B models with Ollama. For 30B and larger models you want 64GB or more of fast unified memory, like the Ryzen AI Max+ 395 or Apple M-series chips.
How much RAM should a mini PC for Ollama have?
32GB is the sweet spot for 8B to 14B models with room to spare. 16GB works for 8B models. 64GB to 128GB is for 30B to 70B models.
Is the integrated GPU used by Ollama on a mini PC?
On Apple silicon, yes, fully. On AMD mini PCs it depends on driver and ROCm or Vulkan support for that chip; many 780M systems end up running on CPU. Check ollama ps to see.
Is a Mac mini good for Ollama?
Yes, it is one of the easiest setups. Get at least 24GB of unified memory, since memory cannot be upgraded later.