Which Mac mini, for which models?
| Configuration | Memory bandwidth | Apple US price | What fits comfortably | AdminOnsite plan |
|---|---|---|---|---|
| M6, 16GB, 256GB SSD | 153GB/s | $899 | 8B models (about 5GB at 4-bit) | — |
| M6, 24GB, 256GB SSD | 170GB/s | $1,099 | 12–14B models (8–9GB) | Practice |
| M6, 32GB, 512GB SSD | 170GB/s | $1,499 | Up to ~20B (gpt-oss:20b is 14GB); 27B is tight | Office |
| M5 Pro (15-core CPU, 16-core GPU), 24GB, 512GB SSD | 307GB/s | $1,699 | Same models as 24GB M6, faster | — |
| M5 Pro (15-core CPU, 16-core GPU), 64GB, 512GB SSD | 307GB/s | $2,699 | 70B models at 4-bit (llama3.1:70b is 43GB) | Firm |
Configurations and prices from Apple’s US store and Mac mini specs page, checked November 13, 2026; model sizes from ollama.com/library. Fit guidance is ours. The M4 Mac mini is no longer in Apple’s US lineup.
Why a Mac mini works for this
Apple silicon uses unified memory shared by the CPU and GPU, so the GPU can work with most of the machine’s memory. That’s what lets a small desktop hold models that would need a large graphics card on a PC. Apple’s 2026 Mac mini adds Neural Accelerators in each GPU core; Apple says the M6 model delivers up to 4.8x faster LLM prompt processing in LM Studio than the M4 Mac mini. Apple describes the M5 Pro model as quiet and efficient, and pitches the line for always-on use.
Set it up as an always-on server
- Install a model serverOllama, LM Studio or Apple’s MLX. Ollama runs as a background service with a local API.
- Pull a model that fitsPick from the table above, leaving headroom for documents.
- Keep the model loadedOllama unloads a model after 5 idle minutes by default; set a longer keep-alive so the first request of the morning isn’t slow.
- Stop it sleepingSet the Mac never to sleep, so the server stays up overnight.
- Add a chat window and document indexOpen WebUI or AnythingLLM for staff, pointed at the folders it should search.
- Add battery backupA small UPS rides out short power cuts.
Questions
Can a Mac mini run an LLM?
Yes. A 16GB Mac mini runs 8-billion-parameter models; 24–32GB runs 12- to 20-billion; the 64GB M5 Pro model runs 70-billion at 4-bit.
Is 16GB enough for local AI?
For 8B models and light use, yes. For an office working on long documents, 24GB or 32GB leaves needed headroom.
Which Mac mini does each AdminOnsite plan use?
Practice: M6 with 24GB memory and a 256GB SSD. Office: M6 with 32GB and 512GB. Firm: M5 Pro with 64GB and 512GB. All are standard Apple configurations, $1,099, $1,499 and $2,699 at Apple’s US list price in November 2026.
Ollama or MLX on a Mac mini?
Both use the Apple GPU. Ollama is simpler to run as a shared server; MLX is Apple’s own framework. Test both with the same model.
Does it need to be on all the time?
For an office server, yes: set it never to sleep and keep the model loaded.
Can I buy a Mac mini and do this myself?
Yes, if someone can maintain it. AdminOnsite is the managed version: configured, updated and supported, yours after 24 months.