Your AI can work
the night shift.

Most of human knowledge, as a download.

Two years ago serious AI meant a data centre and an API key. Today you download a file, and the file writes, reads, listens, speaks, translates and looks at pictures. It runs on a graphics card of €500 to €1,500, in a machine that is neither loud nor hot. This market moves so fast that we struggle to keep up, and we do this for a living.

What already runs on a small local machine:

Language and vision.

Open models such as Qwen read, write, summarise and reason in over a hundred languages, and look at images and scanned pages. Translation between your languages comes with it.

Image generation.

Product pictures, illustrations and mock-ups from a text prompt, on your own card, with no per-image fee.

Speech in both directions.

Whisper turns recordings and calls into text. Kokoro turns text into a natural voice. Both fit on a small machine.

Decisions in milliseconds.

A new class of small models answers a fixed question about a document in one pass instead of writing an answer word by word. Laya, released this month, sorts a ticket or scores an email in about 30 milliseconds. Fast, cheap, and made for the boring work.

And the files keep shrinking. A model that needed a rack of servers two years ago is now a download of about 6 GB, smaller than a film in HD, and it fits in the memory of an ordinary machine.

Fast enough by breakfast.

A chat needs a fast answer. A pile of yesterday’s enquiries does not. A local model can work through the night, sort messages and prepare research while nobody is waiting.

Even a machine without a separate graphics card can do this on its built-in graphics. Slower, but steady, and still useful.

As long as the work is done by morning, speed is not the question. What matters is that it is reliable, that the results are useful, and what it costs to keep running. A machine you own can take its time. Nobody charges you per word.

Give it the jobs you keep putting off.

An inbox that arrives sorted.

Classify incoming mail, separate invoices from enquiries, suggest labels and draft summaries for review.

Customer research, done overnight.

Look up prospective customers from public company information, match them to your criteria and prepare notes with sources for a person to check.

Documents you can find.

Let the model read scans and photos, suggest what each document is and pull out the fields you need, ready for a person to check.

Audio that becomes useful.

Whisper can turn recordings into text locally. Kokoro goes the other way: compact, local text-to-speech.

The tools for all of this are in the catalogue. Connecting them to your inbox, your CRM and your archive is still your own plumbing, or ours together in a pilot. Start with one repeatable job, test it on your own examples and keep a person in the loop where mistakes matter.

Your documents should stay in the building.

Invoices, customer notes and internal documents can be processed on your own machine. Nothing is sent to an AI provider. That alone is a reason to own the machine, even when a cloud model would answer faster.

You still decide where the rest goes. Web research talks to the outside world; mail providers, connected tools and backups each move data somewhere. Know where.

The server you own gets another job.

OurGrove’s beta catalogue already includes OpenWebUI with Ollama, ComfyUI and Kokoro. Our job is to run local AI on the same machine, with the same setup and recovery, as your web apps and files.

The server that hosts your software can now also do part of the work in it. That is a good reason to own it.

Which model have you run locally? Tell us what it did for you ↗