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

# Local model

> Run an assistant entirely on your device, with no account and nothing leaving your machine.

You don't need an external assistant to get AI replies during a call. StageWhisper can download a model and run it on your own Mac, so your calls get answered on-device. No account, no API key, and nothing leaves your machine.

Reach for it when you want a private option that works offline, or a backup for when the assistant you normally connect to isn't around.

## When the local model answers

* **As a fallback.** If you have an external assistant connected ([OpenClaw](/agents/openclaw) or [Hermes](/agents/hermes)), that assistant stays in charge. The local model only steps in when the assistant is offline or not set up, then hands back once it's available again.
* **As your default.** Flip one toggle and the local model answers everything, even with an external assistant connected.

## Set it up

<Steps>
  <Step title="Open Settings → Local Model">
    You'll see two recommended models, a field for any Hugging Face model, and an option to point at model files you already have.
  </Step>

  <Step title="Get a model onto your machine">
    Download the recommended model, paste a Hugging Face repo id, or choose a folder you already have. The download runs in the background, so you can keep working while it finishes.
  </Step>

  <Step title="Wait for Installed">
    Once the model shows Installed, it's ready to answer.
  </Step>

  <Step title="Optional: make it your default">
    Turn on "Use the local model as my default" if you want it to handle every reply. Leave it off to keep it as a fallback behind your external assistant.
  </Step>
</Steps>

## Choosing a model

**The recommended models** come in two sizes, both ungated, so there's nothing to sign up for and no license to accept. The smaller one is a light download that runs on a modest machine. The bigger one wants more memory but tends to answer better, so reach for it if your machine has the room.

**Any Hugging Face model.** Paste a GGUF repo id (for example `unsloth/Qwen3.5-4B-GGUF`) and download it. For gated or private repos, add an access token in the same panel.

**A model you already have.** If you've already downloaded GGUF files, whether with the Hugging Face CLI, another local app, or an earlier StageWhisper download, choose **Use a model from your computer** and point at the folder. Nothing downloads and no token is needed.

<Note>
  Models must be in GGUF format. Pick a folder that has a `.gguf` file inside, or a Hugging Face repo that ships GGUF files. Models kept in Ollama's own format will not work here.
</Note>

## Privacy

When the local model answers, the reply is generated on your machine. Your transcript isn't sent to us or to any external assistant. The model runs through an open-source on-device engine, so all the work happens locally on your GPU.

See [privacy](/guides/privacy) for the full breakdown of what does and doesn't leave your computer.

## Requirements

* macOS 14 (Sonoma) or later, Apple Silicon.
* A few GB of free disk and enough memory for the model you pick. The smaller recommended model is light on both; the bigger one needs more room.
* No account or token, and no network once the model is downloaded.

## Prefer to connect your own assistant?

The local model is just one option. You can also bring a full assistant that already has your memory and tools.

<Columns cols={2}>
  <Card title="Connect OpenClaw" icon="plug" href="/agents/openclaw">
    Pair an OpenClaw assistant and let it act during the conversation.
  </Card>

  <Card title="Connect Hermes" icon="robot" href="/agents/hermes">
    Pair a Hermes agent to follow your call and respond in real time.
  </Card>
</Columns>