INFRASTRUCTURE / LOCAL AI
Install local AI.
And be able to use it tomorrow.
Downloading a model is one step. A business also needs simple access, suitable performance and someone who can maintain the whole thing.
Written by Saphan · 7 September 2026
Define what really needs to stay local
The model can run on your own computer, on a server on the network or on a remote service. The visible interface is not enough to determine the path the data takes. Documents, text recognition, search, embeddings, logs and backups may each follow different paths.
The Ollama FAQ distinguishes between models run locally and its cloud features. Software installed on your computer therefore does not, on its own, mean that all processing stays there. The scope must be verified on the configuration chosen.
Choose a first task
Start with an identifiable output: sorting requests, preparing a short summary or finding a passage in a document. A fictional or authorized corpus is enough to observe the quality, the waiting time and the limits of the hardware. A general conversation is an incomplete test if the final use involves long documents.
Assemble the complete chain
- The engine: a model whose license and format fit the use.
- The computing power: memory, acceleration, storage and headroom for simultaneous requests.
- The interface: access suited to the users, with the accounts needed.
- The data: authorized sources, updates and viewing rights.
- Operations: startup, monitoring, backup and restore.
Opening the service up to the whole company or to the internet is a decision separate from installing it. Prefer access limited to the intended scope, then test the rights from an ordinary user account.
Check the independence you expect
On a test environment, once the necessary files are available, simulate the loss of the external connection. What still works? What stops? Does the system report the absence of a remote source? Repeat after a restart: a model still loaded in memory can hide a dependency at startup.
This test is not, on its own, a confidentiality audit. It lets you check one concrete property: availability. Saphan can prepare the deployment and the first steps in using it, according to your constraints, with or without additional cloud capacity.