Bringing the Model Home
Building My Personal AI, Step 2
At the end of Step 1, I wrote that the room was ready.
This week, someone moved in.
A model now runs on the Mac Studio in my office. The iMac and the MacBook can ask it questions from anywhere in the house, or from anywhere I travel. Nothing goes to an outside service.
It took one script, about 30 GB of downloads, and a few surprises.
The first question
The first thing I asked it was small. "Say hello in five words."
Hello there, how are you?
An answer from a machine you can see. No account. No subscription. No data leaving the room.
The numbers
Before I started, I read the guides. For my model on a machine like mine, they estimated 40 to 70 tokens per second. A token is roughly three-quarters of a word.
So I measured.
| Asked from | Model | Writes | Full answer in |
|---|---|---|---|
| Mac Studio | Qwen 3.6 (main) | ~130 tokens/sec | 3.1 s* |
| Mac Studio | Gemma 4 (fast) | ~94 tokens/sec | 0.6 s |
| iMac, over Tailscale | Qwen 3.6 (main) | ~100 tokens/sec | 1.1 s |
| MacBook, over Tailscale | Qwen 3.6 (main) | ~104 tokens/sec | 1.1 s |
| The guides' estimate | Qwen 3.6 (main) | 40–70 tokens/sec | — |
* The first time it saw this question, Qwen spent about 2 seconds reading it. Later runs from the iMac and MacBook reused that work, so they finished faster.
About 130 tokens a second on the Studio. Roughly double the estimate.
For scale, people read at about four or five words a second. This writes about a hundred.
From the MacBook, over the private network, a full answer arrived in about one second. The network cost almost nothing.
What surprised me
The bigger model was faster.
I run two models. A large one for depth and a small one for speed. I expected the small one to win on speed.
It didn't.
The large model is built as a "mixture of experts." It holds 35 billion parameters, but only about 3 billion work on any given word. Think of a large team where only the right few specialists speak up for each question.
So it carries the knowledge of a big model, at the speed of a small one.
That was the reason I chose it. Seeing it in the numbers was still a small thrill.
The first try is always slower.
The first time I asked, the model had to load into memory. Ten seconds. After that, it stays loaded for half an hour, and answers come in about a second.
One run from the MacBook was half as fast as the next. Nothing was wrong. The Studio was simply busy for a moment.
The lesson I keep relearning: measure twice before believing either number.
My own test lied to me.
My first speed test had a column for how fast the model reads your question. On the second run it reported 1,290,323 tokens per second.
That's not a fast model. That's a broken measurement. The server had remembered the question from the first run, so it barely had to read anything, and my math divided by almost zero.
I replaced it with something simpler: how long you wait before the answer starts. That's what a person actually feels.
It's easy to measure what's convenient instead of what matters. I did it within the first hour.
The first real decision
The model server has no password.
Right now, any device on my home Wi-Fi could use it. Not just my Macs. My phone. The TV. A guest's laptop.
It isn't reachable from the internet. But inside the house, it's open.
For now, with just me, that seems an acceptable trade-off. The simple fix until then: guests go on a separate Wi-Fi.
But it made something clear. The moment this becomes a household AI, the questions change. Who is allowed to use it? Who can see what? Who did what, and when?
Those aren't features. They're the foundation. That's Phase 2.
What it feels like
I wrote at the start of this series that I wanted to know what it would feel like to ask a question and know the answer never left the house.
Maybe the test isn't whether it feels revolutionary. The personal computer didn't, once it was on your desk. It just felt like yours.
The how
The scripts, the settings, every number, and what went wrong:
Next: Step 3. Giving it a memory.