Who this is for
Private Local AI is built for small and medium organisations in the Perpinyà – Girona – Barcelona corridor who want their own model, on their own hardware, with a person nearby who can be in the room. Not a datacentre promise. A machine on your desk and someone accountable for it.
That is a deliberate limit and a deliberate advantage. We work within a day’s drive, so "local" is a claim we can keep: the machine and the support are both in the corridor. When something drifts, the fix is a visit, not a ticket queue in another timezone.
The mechanism, because it is what makes the price work
The economics come from three mechanisms, not from luck. Weights stream from ordinary system DDR and NVMe instead of being pinned in expensive VRAM, so the memory you already buy is the upgrade path. A Mixture of Experts keeps resident only the experts a query actually uses, so a large model behaves like a small one on the tokens it does not need. Per-tensor quantisation writes the bulk of the weights narrow while protecting the sensitive tensors, so more model fits on the card that is already in the box.
Each of those is measured, not asserted: one model file, identical weights and prompt length across engines, read and write timed apart, versions recorded beside the figure. The benchmark index is the citation, and we will run it on your rig before we recommend anything.
Compliance is topology
The reason a local deployment clears a compliance review without a fight is not a clause. It is that inference happens inside your building. Personal data never leaves the premises, so there is no Chapter V transfer to arrange, no standard contractual clauses to keep valid, no adequacy argument to re-litigate, no foreign disclosure request to answer for. The violation is not promised against; it is never committed, because the data does not go anywhere that could trigger it.
That is an architecture fact, not a policy page, and it is the same for a clinic, a law office or a small manufacturer. Their documents answer questions from a model on their network, and the documents do not move.
Start with the machine you own
The first step is not a purchase. It is an on-site assessment on your own box: we measure what it already does, show you the numbers, and tell you honestly where the limits are. If the machine you have is enough, that is the recommendation. If it is not, you will know exactly what to add and why.
Write to privatelocalai@proton.me and ask about an on-site assessment. We serve the corridor, in person.
The measurements behind every figure we quote: https://localai.hex7.com/benchmark.html.
Leave a Reply