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

# Work offline with local models

> Keep your voice on your computer with the local-only path

SpeechKit can run without cloud keys. On the local path your voice is
processed on your computer. This guide shows how to set that up on Windows.

## What you need

* A Windows 10 or 11 computer with the installed app. See
  [Install on Windows](/speechkit/install-windows).
* One internet connection for the first setup. The installer includes the
  local speech engine. The speech model itself is downloaded when you choose it.

## Steps

<Steps>
  <Step title="Choose Local in the setup">
    In the first-run setup, choose **Local**. It runs Dictation and Assist on
    this device, and the models live on your computer.
  </Step>

  <Step title="Pick a speech model">
    Pick a local Whisper model to download. If you do not choose one, SpeechKit
    starts with the smallest starter model. The download runs in the
    background, and you can continue while it finishes.
  </Step>

  <Step title="Test your microphone">
    The setup can record a short local test. It is optional, and nothing is
    saved. The text is shown only in the setup window.
  </Step>

  <Step title="Set the privacy scope">
    Open **Settings > Privacy & Data**. Under **Network scope** choose **This
    device only**. Only models that SpeechKit runs on the device may then be
    used. Everything that needs the network is paused, not deleted.
  </Step>

  <Step title="Keep nothing you do not need">
    Choose **Zero retention** if raw meeting material should be removed after
    a meeting. See [Choose your privacy settings](/speechkit/privacy-settings).
  </Step>
</Steps>

## Good to know

* With **This device only**, even model downloads and update checks are
  blocked by default. If you need to download a model later, widen the scope
  for that moment. A setting called `allow_setup_traffic` in `config.toml`
  can allow model and update downloads on official sources, without audio,
  transcripts or prompts.
* Text you dictate stays in your history on the computer. A computer without
  internet still saves transcripts unless you choose Zero retention.
* Assist and the Voice Agent can also run on local models. A local language
  model and local speech output are separate downloads or services. Check
  the **Model** tab of each mode.
* On macOS, the Dictation beta works with the bundled local engine, and models
  download on demand.

## If something goes wrong

* **The model does not download.** With a strict scope, downloads are blocked.
  Widen the scope for the download, then switch back. Also check free disk
  space.
* **The model download stays stuck.** Retry the download in Settings.
* **Dictation says the engine is not ready.** The local engine may still be
  starting or the model is missing. Look under **Needs attention** in
  **Overview**.
* **A setting is greyed out.** It needs the network. Settings shows why it is
  disabled.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.