Ask what a small AI appliance can do and you will usually be handed numbers: memory, throughput, a benchmark. They are precise and almost useless, because none of them tells you whether the thing will answer your question.
Here is the same subject in terms you can actually judge.
How long a document
A machine of this size handles ordinary documents comfortably: letters, contracts, reports, statements, the things that arrive in a working week. It can read a long document end to end and answer questions about it.
Where it strains is holding several very long documents in mind at once — comparing four full contracts clause by clause in a single pass, for instance. That work still happens; it just happens in stages rather than all at once, and it takes longer.
How large an archive
This is the question people expect to be limiting, and it is the one that is not.
The size of the archive a machine can draw on is not really constrained by the machine. Years of correspondence, decades of paperwork, tens of thousands of documents — these are storage and organisation problems rather than capability problems. A small appliance can be the front end to a very large archive.
What the archive's size does affect is preparation time. A larger archive takes longer to get into a usable state, which is work done once rather than a permanent limit.
How fast an answer
Conversational, not instant. A straightforward question comes back in seconds. A question that requires reading across many documents takes longer — sometimes tens of seconds — because it really is reading.
This is slower than a large hosted service, and it is worth being honest that you will notice. In practice it matters less than expected, because the questions worth asking your own archive are not the ones you fire off while waiting for a kettle.
What it is genuinely good at
Anything grounded in your own material. What something cost, what a document says, what changed between two versions, who has not replied. This is its home ground and it is better here than a far larger model with no access to your papers, because the larger model is guessing.
Routine language work. Drafting, summarising, restructuring, turning notes into something readable.
Noticing. Patterns across documents — a duplicated invoice, a renewal that rose, an annual letter that stopped arriving.
What it is not good at
General knowledge at the frontier. Obscure history, specialist science, an unfamiliar language. A large hosted model is better and will stay better, and the sensible arrangement is to let you send such a question out when you decide it is worth it.
Very long chains of reasoning. Multi-step problems where each step depends on the last. It can do them; it is less reliable than the largest models, and the failures compound.
Anything requiring the newest information. A machine in your building knows your material and what it was trained on. It does not know this morning's news unless something brings it in.
Why we describe it this way
We deliberately do not publish model names or benchmark figures, for a reason that is practical rather than coy.
Those numbers change every few months. A page built on them is out of date almost immediately, and worse, it invites you to compare the wrong thing — whose model is larger — when the question that actually determines whether this is useful to you is whether it can read your material at all.
What we run gets better over time, and improvements arrive as part of the service rather than as a purchase. The capability described above is the floor, not a ceiling.
The honest summary
A small machine will not win an argument about raw capability with a data centre. It is not trying to.
It is trying to be the only thing in the room that has read your papers — and on every question that depends on that, it is not close.
What can and cannot be trained on →
Local AI. Private data. Local training.
$8,995 the first year — everything included. $4,995 each year after. The first year costs more because it contains the Jetson Orin placed and configured, your archive loaded and the first training run; every year after is the service running.
Or add the Archive Assessment first — $1,950, credited in full
You order and pay at checkout; we write back within days — a decline returns every dollar, and until our letter confirms the year you may withdraw in writing. From that letter the year is final, and it arrives on the date the letter names. Prefer to write to us first?