Local AI · small business · Mac mini
Local AI on a Mac mini for a small business
Measured on our own 16 GB Mac mini in Atlanta.
Staffbox runs an open-weight model on a 16 GB Mac mini in your building, through Ollama. It drafts quotes from your price list and answers questions from your own files. Your data stays on your network; cloud AI is off by default and only runs on your own key.
Send 3 requests, freeOr call Hadi: 404 493 6248
No customer install yet. Scores come from fictional test companies.
- Machine
- Mac mini M4, 16 GB
- Model
- qwen3:8b through Ollama, listening on 127.0.0.1 only
- Score
- 351 of 360 generated messy requests, 0 wrong totals in 475 answers (6 Oct)
- Speed
- about 1 to 3.2 s average per answer, one request at a time
- Power
- 4 W idle, 65 W max (Apple's rating)
- Cloud
- off by default; on only with your key, in writing
What the 16 GB Mac mini runs
One model, qwen3:8b, kept warm in memory, with your company notes (the brain) and the quote action. It answers one request at a time; requests that arrive together wait in a queue.
| Test set | Right | Average | 90% within |
|---|---|---|---|
| Fieldstone IT | 40 of 40 | 2.3 s | 5.3 s |
| Peachtree Cabinet Works | 39 of 40 | 2.5 s | 4.4 s |
| Messy requests, set 1 (seen while fixing) | 19 of 20 | 1.2 s | 2.3 s |
| Messy requests, set 2 (never used while fixing) | 15 of 15 | 1.1 s | 1.7 s |
| Generated, Fieldstone | 201 of 204 | 1.6 s | 2.1 s |
| Generated, Peachtree | 150 of 156 | 3.2 s | 3.4 s |
Scroll the table sideways.
Quote action v2.4, 6 Oct 2026, 16 GB Mac mini M4, one request at a time, model warm. measured Every graded answer: the Staffbox scorecard on Hugging Face.
What does not fit: a 14B model timed out on the 16 GB box, so we don't use one there. The gpt-oss-20b option, for buyers who cannot use Alibaba's Qwen models, needs 32 GB; on our GPU test box it scored 39 of 40, 40 of 40 and 15 of 15, and its speed on Apple hardware is not yet measured. Pilots run on a 32 GB unit, measured on the same tests before day 0. All options side by side: local AI models and their scores.
Power draw
Apple rates the Mac mini (M4) at 4 W idle and 65 W at maximum (Apple: Mac mini power consumption). We have not measured our own unit's draw.
Your data stays on your network
- The model runs through Ollama on 127.0.0.1, so only programs on the box itself can reach it.
- No cloud fallback is configured. Cloud AI runs only if you ask in writing and supply your own key.
- The installer turns off terminal, code execution, browser, computer use, web and delegation tools.
- No one trains a model on your data.
- Staffbox has no sending code; a person sends.
Every control and its status: the Staffbox Trust Center. What it does with RFQs: RFQ automation from your own price list.
Bigger than a Mac mini: Mac Studio, planned
A Mac Studio is planned, not yet measured. The model class it would run, a 27B model, scored 40 of 40 and 39 of 40 with the brain on our GPU test box on 30 September. We will publish Studio speeds once we have measured them. Your scorecard decides the box: we move a workflow to a bigger one only when the test shows the smaller one cannot pass it.
Start free: 3 requests, then 30 past quotes
Step 1 · free
3 requests
Send three requests your team answers every week. Hadi sends back what Staffbox would have drafted, misses included. No card, no commitment.
Step 2 · free
30 past quotes
Send 30 past quote requests and your price list. We run them on our Mac mini in Atlanta, your team grades 10 answers blind, and we delete the files afterwards. Nothing is installed for this step.
If you go ahead: founding sites (first 100) get a free 30-day pilot, then $595 a month for as long as the box is live, no onboarding fee. draft offer
Questions
Can a Mac mini really run AI for a business?
For one focused job, yes in our tests: a 16 GB Mac mini with qwen3:8b drafted quotes for two fictional companies at 40 of 40 and 39 of 40, about 1 to 3.2 seconds average per answer. It handles one request at a time; others queue.
Does any of our data go to the cloud?
Not by default. The model runs on the box, and cloud AI is off unless you ask in writing and use your own key.
How much power does it use?
Apple rates the Mac mini (M4) at 4 W idle and 65 W at maximum.
What if we need a bigger model?
The gpt-oss-20b option needs 32 GB, and a Mac Studio is planned but not yet measured. We move a workflow to a bigger box only when the test shows the smaller one cannot pass it.