Our own AI model on a Mac mini: the build log.
We call it the Engine. What we are building, why it starts on a Mac mini, what runs on it first, and a dated log of what happens next. Started 5 November 2026.
The short answer
Sentry AI is building its own open weight model stack, starting on a Mac mini in its Auckland office. Its first jobs are running our own internal operations and our daily eval pipeline. The point is proof: that a business can run a capable AI model on hardware it owns, in New Zealand, with its data never leaving the building. We are doing it in public, one dated log entry at a time, and we publish a number only once we have measured it.
What we are building, and why
Most of our work is moving teams onto Claude or ChatGPT, and for most businesses a hosted frontier model is the right answer. The Engine is the longer plan behind that work: an open weight model, running on a machine we own, doing a real share of a real business's daily AI work. Three reasons make it worth building rather than reading about.
Sovereignty
A model on a machine in our own building answers the question every regulated client asks first: where does the data go? Nowhere. Prompts, documents and answers stay on the desk they started on.
Cost per query
An API meters every token forever. A machine is paid for once. Which is cheaper depends on volume and on how often the small model gets the job right, so we will measure cost per completed task on our own work rather than quote anyone else's number.
The repricing risk
Every business running on a hosted model has a supplier who can change the price, the model or the terms. We want to know, from our own experience, how much of a business's daily AI work can run on a model nobody can reprice or switch off.
The wider argument, where a model runs, where the data rests and who can switch it off, is on our sovereign AI page.
The hardware we start on: a Mac mini
We start on a Mac mini because it is the kind of machine a business can buy, plug in and keep on a desk. No server room, no rack, no new power circuit. If a model on a desk can carry a useful share of our work, that result transfers to a law firm, a medical practice or an accounting firm in a way a cluster of GPUs never would.
The one constraint that matters on a machine like this is memory: the model and the conversation have to fit, or it slows to a crawl. That is why the smaller sizes of the open weight families exist, and why the open weight models page spends so long on the question of when a small model is enough. Which model we run, and in what size, goes in the log on the day it is running.
What runs on it first
We are our own first client. Sentry AI runs its business on its own second brain, and the agents that read and write it do a lot of narrow, repeated work. That work moves to the Engine first, where a wrong answer costs us and nobody else.
Our own operations
The repeated, checkable work inside Sentry AI: sorting, summarising, drafting from our own templates, answering questions from our own pages. The kind of work the open weight models page says a small model is usually fine for.
The daily eval pipeline
Every morning an eval reads what our AI did the day before and picks at most one thing to fix. Running that judgement on our own hardware is the first serious test, because it is real work with a real consequence when it is wrong.
How the daily judgement works, and why it reads what was done rather than what was typed, is on the AI evals page.
Dated, newest first
The build log
An entry is added when something happens, with the date. No entry means nothing has changed since the last one.
Page created; the build log starts here.
- The model running on it, named, with the date it went live
- The jobs it took over from a hosted model, and the ones it could not
- Cost per completed task against the API, worked out the way the hub page describes
- What broke, and what we changed
- Where it sits once it outgrows the desk
What it means for your business
Building towards, not selling today
On-premises custom models are what we are building towards. We do not sell them today, and this page is not an offer. If your data must stay in your control or in New Zealand, the options you can act on now are a frontier model pinned to an Australian region, or private LLM infrastructure on hardware you own. We will tell you honestly which fits, and whether a business plan on Claude or ChatGPT is simply enough.
What this build gives you in the meantime is a record you can check: what a small machine in an office actually managed, on real work, with the failures left in.
Frequently asked questions
What is the Engine?
The Engine is Sentry AI's own open-weight model stack, run on hardware we own. It starts on a Mac mini in our Auckland office, and its first job is running our own internal operations and our daily eval pipeline. It is a build we are doing in public, not a product.
Why start on a Mac mini rather than a server?
Because the point is to show what a business can do on a machine it can buy, plug in and keep in its own building. A Mac mini is small, quiet, and sits on a desk. If the work we run on it holds up, that is a more useful result for a New Zealand business than a benchmark from a rack of GPUs would be.
Which model is running on it?
We will name the model on this page when it is running, with the date. Open weight families such as Qwen3, gpt-oss and Gemma 4 come in sizes small enough for one machine, and our open weight models page explains how we compare them. Until an entry in the log says otherwise, nothing is running yet.
Can we buy one of these from Sentry AI?
Not today. On-premises custom models are what we are building towards, and the Engine is how we earn the right to offer them. If your business has data that must stay in your control or in New Zealand, the conversation we can have now is about Claude pinned to an Australian region, or private LLM infrastructure, and we will tell you honestly which fits.
Does the Engine run the AI that your clients use?
No. Client teams run on Claude or ChatGPT on their own accounts, with their data handled as each engagement sets out. The Engine runs Sentry AI's own operations first. Anything that reaches a client goes through the same measured rollout as every other piece of work we do.
How will we know the results are real?
Each log entry is dated and says what was done. When we publish a comparison it will say which model, which tasks, how many runs, and how it was scored: whether the task was completed, how long it took and what each completed task cost. Numbers appear on this page only after they exist.
Wondering whether your business needs its own model?
Tell us the job and the data. We will say plainly whether a hosted model is enough, and what to watch on this page if it is not.
Ask an AI about this page
Opens the assistant with this page loaded: read it, summarise it, cite it.