01 / How it works
How a private local AI test works
Test an open model on the hardware you are considering before you buy it.
Choose a setup.
Pick the hardware and the model from a short menu of setups we have prepared and tested ourselves: an NVIDIA DGX Spark with 128 GB of memory, running Qwen, DeepSeek or GLM. Every setup names its exact machine and pins its model to a version.
Connect your own tools.
Each setup is up and running and answering seconds after you start. You receive a web address and a key for an AI-compatible endpoint; any tool that accepts those two things works, and nothing is installed on your network.
Run it and decide.
Run the tasks you care about and note the speed, the quality and the errors. Then buy, do not buy, or ask us for help. If the answer is no, you have saved the purchase price. If it is yes, you keep the exact configuration, so it can be set up again on a machine of your own.
Private Local AI in practice at a Law Firm
Professional secrecy means our client data cannot be shared with OpenAI or Anthropic. Also, we don't want our IP to end up there.
Running our own AI is hard, and we could not tell whether it would handle all our work.
Private-Local.ai lets us run our real workflows on the exact hardware and AI models we were considering. Within an afternoon we knew on what setup we would bring AI into the firm on a machine and model we own.
Mina
Partner
02 / Questions
Private AI questions, answered.
What is private AI?
Private AI means running AI models on hardware your organisation controls, so the work stays inside instead of going to an outside provider. The test drive is the step before it: you rent the machine by the hour, on hardware that is not yours, to find out whether owning one would do your work well enough.
Where are the servers?
The servers for the United Kingdom are in London. Each country site names the location of its own servers: United States (New York and San Francisco), United Kingdom (London), Germany (Frankfurt am Main), France (Paris), Spain (Madrid), Italy (Milan) and Brazil (São Paulo).
Why not just keep using ChatGPT or Claude?
For much of your work, you should. ChatGPT, Claude and Gemini give better answers than any open model small enough to run on a desktop machine. A machine of your own pays off when you use AI heavily or need control over where your data goes, not because its answers are better.
Which tools can we connect?
Anything that speaks to an AI-compatible endpoint, which today is most tools: coding assistants such as Continue, Cline and OpenCode, automation platforms such as n8n, note-taking tools such as Obsidian with the Copilot plugin, and your own code. So far we have tested the Hermes Agent and direct calls from code ourselves; the others accept such an endpoint in their settings, but we have not tested them yet. If you are unsure about yours, tell us which one it is.
What does it cost, and why not rent a bare machine?
£10 an hour, and you pay only for the hours you use; most questions are answered in one or two. A bare DGX Spark rents for much less, but you install and serve the model yourself. Our rate pays for a machine that is prepared, configured as disclosed and answering before you start.
What happens after the test drive?
You decide with your own results in hand: buy nothing, buy a machine and set it up yourself, or ask us to help with that as separately scoped work.
Dedicated hardware
Request a hardware test drive.
Run your own tasks on a dedicated machine with the NVIDIA GB10 chip of the DGX Spark before you buy one. Hardware test drives are offered on request; tell us about your workload through the contact form.
Request a hardware test drive