## Where this stands

**On-Premise Brain is not generally available.** Today Brain runs
Levirge-hosted and tenant-isolated, and self-hosting is a conversation we
have per organisation rather than a package you can download. This section
exists so you can tell us you need it, and so the shape of that conversation
is public rather than hidden behind a sales call.

If that is a blocker for you, the honest answer is: register below, and talk
to us about timing.

## What is already private

Worth separating from deployment, because it is the reason many people ask
for on-premise in the first place:

- **Hosted model lane** — categorisation, summarisation and chat run on
  models Levirge hosts on its own infrastructure, with no cloud model
  provider in the path. Embeddings and reranking run on Levirge
  infrastructure in both lanes.
- **Vault isolation** — knowledge is partitioned per project, product or
  client, and Personal Vaults fail closed.
- **Tokens you control** — per-user workspace tokens you issue and revoke.

That covers "no third-party model sees our data". It does not cover "the data
never leaves our network" — that is what on-premise is for.

## The two questions

- [Requirements](/brain/docs/on-premise/requirements) — hardware by what
  you already have, the services around it, and what sizes a deployment.
- [Model options](/brain/docs/on-premise/model-options) — the four inference
  models, and where each one can run.

The short version: **Brain self-hosts on commodity hardware.** Search is
equally fast with or without a GPU, and the one component that genuinely
wants a discrete GPU — reranking — can be turned off, with search falling
back to rank fusion.