# Choose a smart search model

Source: https://frameleaf.app/docs/using/search-models/

The smart search model decides how well Frameleaf understands your words and your photos. The one Frameleaf starts with works well for most libraries. Change it when you want quicker search on a small server, better results on stronger hardware, or search in more languages.

This page is for administrators. Changing the model means every photo and video is read again, and searching by description finds nothing until that finishes. Searching by date, place, person and file name keeps working, and your albums, tags and captions are not touched.

## Where to find it

Open **Settings**, then **Search & intelligence**, then **Visual search**. The **Search quality** slider shows the models Frameleaf publishes.

[![The Visual search settings page: Visual search is on, the Search by describing a photo switch, and the Search quality slider with Good, Better and Best stops and a Frameleaf Cloud range, with the chosen model and its estimated time below.](https://frameleaf.app/shots/docs/search-quality-2026-11.jpg)](https://frameleaf.app/shots/docs/search-quality-2026-11.jpg)

| Model on the slider | Good for |
| --- | --- |
| `ViT-B-32__openai` | The quickest smart search, for small or slow servers |
| `ViT-B-16-SigLIP-384__webli` | Accurate smart search for most libraries. The default. |
| `ViT-SO400M-16-SigLIP2-384__webli` | The best results. It needs more memory. |

The colours on the slider show where each model can run: green on your graphics card, white on the processor, and blue on Frameleaf Cloud. A blue stop runs the same model on Frameleaf Cloud instead of your own hardware, and you see the cost before anything is sent. If this server isn't linked to a Frameleaf account yet, choosing a blue stop opens the Frameleaf Cloud setup first.

## Switch to another model

1. Open **Settings**, then **Search & intelligence**, then **Visual search**.
2. Move the **Search quality** slider to the model you want.
3. Read the notice that appears. It says how many photos and videos will be read again and about how long that takes at your server's recent pace.
4. Choose **Save and start now**, or **Save and start tonight** to leave the work for the [overnight schedule](https://frameleaf.app/docs/admin/jobs/#nightly-tasks). **Undo** puts the slider back.

The old search results are cleared when you save. Photos can be found by describing them again one by one as each is read, so search fills in as the work goes. Follow the progress in **Settings**, then **Computers & background work**, then **Background work**.

Each model keeps its own search data. If you switch back to an earlier model, the library is read again for that model too.

## Use a model of your own

To search in another language, or to try a model that isn't on the slider:

1. On **Visual search**, open **Advanced: use your own model**.
2. In **Hugging Face model**, type the model's full name, like `organisation/model-name`.
3. Save, and start the reading as above.

Frameleaf checks the model's files and settings before anything is read again. While your own model is in use, the slider shows no stop and a note names the model. Empty the box to go back to one of the published models.

For multilingual search, try a few typical searches in each language once reading finishes, and compare the results on your own photos.

## What else changes

Suggested tags, the built-in smart albums and moments in videos also use what smart search has read, so they catch up as the library is read again. Your originals, your own tags and the descriptions you wrote stay exactly as they are.

## Hardware and where it runs

Smart search runs on your server unless an administrator sends it to another computer at home or to Frameleaf Cloud. Frameleaf Cloud only runs the same model your server uses, so results always match. See [Search and faces on Frameleaf Cloud](https://frameleaf.app/docs/admin/workers/#search-and-faces-on-frameleaf-cloud). Speed it up with [hardware acceleration](https://frameleaf.app/docs/admin/hardware-acceleration/) on a supported graphics card, or run it on another computer with [remote machine learning](https://frameleaf.app/docs/admin/remote-machine-learning/).

Models come from your configured model source, and Frameleaf reports a missing model as a failure so you always know where your models come from. See [Where models come from](https://frameleaf.app/docs/admin/environment-variables/#where-models-come-from).
