# Vision & Audio

Actual can run models that understand images and audio on hardware you control. These capabilities depend on the model you have loaded and on the app you use to send the message.

After a connected app sends an attachment to Actual, inference runs on your Actual computers rather than a third-party hosted model. The connected app may have its own storage or sync features, so check its privacy settings too.

## What You Can Do

With vision, you can attach an image and ask the model to:

- explain a chart, diagram, or screenshot
- summarize a scanned page or receipt
- identify objects and details in a photo
- answer questions about what it sees

With audio, you can attach a recording and ask the model to:

- transcribe speech
- summarize a meeting or voice note
- list decisions and action items
- answer questions about the recording

The quality of the result depends on the model. A clear prompt that says what you want from the attachment usually works best.

## Choose a Compatible Model

**Gemma 4 E2B IT** in the [Actual model registry](https://models.actual.inc) is an option for image and audio input. Download it through Actual so the complete model arrives together, then load it before sending an attachment. Support depends on the installed client and hardware as well as the model, so check the capability icons after loading. Other Gemma 4 variants may support images without audio. [Model Discovery](/docs/model-discovery) walks through downloading and loading models.

On macOS, capability icons appear beside **Model** in the Actual menu after the model finishes loading:

- the eye means image input is supported
- the waveform means audio input is supported

Hover over an icon to see its description. The icons describe the model that is loaded now, so they may change when you switch models.

## Use Vision or Audio

Console Chat currently accepts text only and does not show attachment controls. To send an image or audio clip, use a compatible application that is connected to Actual and supports that kind of attachment. If you still need to connect an application, start with the [API Quickstart](/docs/api-quickstart) or the application's custom-provider instructions.

Once the application is connected:

1. Make sure your vision- or audio-capable model is loaded in Actual.
2. Use the application's attachment control to add a local image or audio clip. Pasting a web link is not the same as attaching the file.
3. Write a prompt that says what you want the model to do with it.
4. Send the message as you normally would.

For example:

- Attach a screenshot and ask, “Read this error and explain what I should try next.”
- Attach a chart and ask, “Summarize the trend and call out the largest change.”
- Attach a recording and ask, “Transcribe this, then list the action items.”

Applications vary: some support images but not audio, even when the loaded model supports both.

## File Types

For the most reliable results, use:

- **Images:** PNG or JPEG
- **Audio:** WAV or MP3

Other image formats may work, but converting to PNG or JPEG is the simplest first step if an image is rejected. Convert unsupported audio to WAV or MP3 before trying again.

Actual accepts up to 16 MiB for the complete message after the app prepares it for sending. This total includes your prompt and attachment. Resize an image or shorten an audio clip if the app reports that the message is too large.

## Troubleshooting

### The capability icon is missing

Wait for the model to finish loading. The eye must be visible for images, and the waveform must be visible for audio. If either required icon is missing, check the installed model and client version before sending an attachment. A model with neither icon may support text only.

### A download was interrupted or the capability still does not appear

Check download progress and errors, then retry the download through Actual if needed. Wait for installation to finish and load the model again. If you restart Actual, expect active work to be interrupted.

### The app accepts an attachment but the model does not respond to it

Confirm that the app supports image or audio messages when connected to a custom provider. Some apps only enable attachments for their built-in services. Also check that the matching eye or waveform icon is visible for the loaded model on macOS.

### A file is rejected

Try a PNG or JPEG image, or a WAV or MP3 audio file. Reduce or shorten the attachment if the complete message is larger than 16 MiB.

### It still does not work

Update Actual with `actual update`, restart it, and load the model again. If Console Chat is the app you are using, its lack of attachment controls is expected; use a compatible connected application for vision and audio.

## Related Docs

- [Model Discovery](/docs/model-discovery)
- [API Quickstart](/docs/api-quickstart)
- [Chat](/docs/chat)
- [Private Relay](/docs/private-relay)