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

# Custom models

> Search Hugging Face from inside SpeakoFlow and download any GGUF model to use as your local assistant, including vision-capable ones. Covers choosing a quantization and what the Include screen vision toggle does.

You can pull any GGUF model from Hugging Face and run it as your local assistant model.

Go to *Settings → Assistant → Browse and download models*, then choose **Browse Hugging Face**.

<Note>
  This is for **assistant** models only. Speech-to-text uses the curated catalog in [Transcription models](/docs/models/transcription-models), and there is no equivalent search for it.
</Note>

<Steps>
  <Step title="Find a repository">
    The dialog opens on the most-downloaded GGUF repositories. Type to search by model or creator. Results are ordered by downloads and capped at 25.

    Each result shows download and like counts, plus a **Vision** badge when the repository's tags mark it as multimodal.
  </Step>

  <Step title="Choose a quantization">
    Open a repository and SpeakoFlow lists its downloadable `.gguf` files, smallest first, with the quantization label pulled out of each filename.

    **Q4\_K\_M** is flagged as recommended. It is usually the best balance of quality, speed, and memory.
  </Step>

  <Step title="Optionally include screen vision">
    If the repository ships a multimodal projector (`mmproj-*.gguf`), an **Include screen vision** toggle appears, on by default. It downloads that extra file alongside the weights so the model can read images, which is what [Screen vision](/docs/assistant/screen-vision) needs.

    The toggle does not appear at all when the repository has no projector.
  </Step>

  <Step title="Download">
    The model is added to your catalog immediately and the download starts in the background, so you can close the dialog and watch progress on the model card.
  </Step>
</Steps>

Downloaded custom models appear under **Your models** in the assistant's model list, and are deletable like any other.

<Warning>
  Custom models are not officially supported. Hardware needs and compatibility vary from model to model, and a model that downloads fine can still fail to load or answer badly.
</Warning>

<AccordionGroup>
  <Accordion title="Which files in a repository are offered">
    Only files at the top level of the repository, since those are the ones that can be downloaded directly. Importance-matrix files are filtered out because they are not loadable models.

    When a repository has several projectors, the higher-precision one is preferred automatically.
  </Accordion>

  <Accordion title="How a custom model is remembered">
    SpeakoFlow records where each one came from (repository, download URL, and projector) in a `custom_models.json` file next to the models. It can re-download one later instead of leaving you with an anonymous `.gguf` on disk.

    Downloads use the same resumable, retrying path as the built-in models.
  </Accordion>
</AccordionGroup>

See [Models and providers](/docs/assistant/providers) for how a local model fits into the assistant.