> ## Documentation Index
> Fetch the complete documentation index at: https://docs.omnia-voice.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Owning your model

> The model trained on your traffic is yours: run it on Omnia, deploy it elsewhere, or take the weights and run it air-gapped — with the proof it works.

When you fine-tune or [self-improve](/reference/self-improvement) a model on
Omnia, the result is **yours**. Unlike closed fine-tuning platforms, where the
model you trained stays locked inside the vendor's serving stack, Omnia gives
you the weights and a choice of where they run.

## Three ways to run it

<CardGroup cols={3}>
  <Card title="On Omnia" icon="server">
    Deploy to a [dedicated endpoint](/dedicated/overview) and call it through
    the same API as any model. Zero ops; billed per GPU-hour.
  </Card>

  <Card title="Another provider" icon="cloud">
    The weights are a standard checkpoint. Run them on any provider that serves
    open-weight models. No Omnia-specific format, no lock-in.
  </Card>

  <Card title="Your own infrastructure" icon="lock">
    Download the weights and serve them yourself: on-premise, in your own
    cloud, or air-gapped. The model, and your data, never leave your control.
  </Card>
</CardGroup>

## Why ownership matters

A fine-tune encodes your proprietary domain logic: your tone, your policies,
your task. Building that into a platform that then holds it hostage behind
premium serving fees is a poor trade. Ownership means:

* **No serving lock-in.** You are never forced to keep paying one vendor's
  compute rates to use a model you trained.
* **Portability.** Standard weights run anywhere open models run.
* **Sovereignty.** For regulated or air-gapped deployments, the model runs
  where your data already lives, with a local, auditable footprint.

## The proof travels with the model

A downloaded model is only as valuable as the evidence it works. Every model
Omnia trains carries its **validation report**: the calibrated judge it was
trained against, its corrected pass rate on your held-out grades, and the
confidence interval, the same [validated-evals](/judges/calibration)
numbers, produced by a judge measured against your own ground truth.

So the deliverable is not just a weights file. It is a weights file **plus the
measured proof** that it clears your bar, the part no benchmark and no vendor
dashboard can give you, because the bar is defined by your grades.

<Note>
  Deployment options depend on the base model and your account's serving
  entitlements. If a model can be trained but not yet served on a given path,
  Omnia surfaces that plainly: you always know which of the three routes is
  available for a given model.
</Note>