# Models

Source: https://help.zira.us/docs/ai-vision/models
Summary: See every model for this data source, send one to a camera, build a camera file, and fine-tune a better version.
Updated: 2026-09-16

A **model** is what lets a camera recognise something — a defect, a part, a label. The **Models**
tab lists every model belonging to this data source and is where you send one to a camera.

![Models tab](/docs/data-sources/models-01.png)

Open a data source, then click **Models**.

## How a model gets here

1. Create and label a dataset — see [Datasets](/docs/ai-vision/datasets).
2. Save the dataset as a version.
3. Run training on that version — see [Training](/docs/ai-vision/training).
4. The finished model appears in this list.
5. Use **Push to Device** to send it to the camera.

You can also use **Add model** (the **+** button) to upload a model you already have, or copy one
from another data source.

## What you see in the list

| Column | What it tells you |
| --- | --- |
| Name | The model's name. |
| Version | Which version of that model this is — 2 means the second one training produced. |
| Format | `PyTorch` is the trained model, `TFLite` and `LiteRT` are the converted copies a camera runs. |
| Looks for | What the model detects. Use the filter icon above the table to narrow the list. |
| Training run | The run that produced it. Click it to open that run in a panel over this tab, so you keep your place in the list. |
| Size | How big the file is. |
| Created | When it was made. |

A dash under **Looks for** just means nobody has set a type yet. You can set one from the row menu.

### Two kinds of file

- A **`.pt`** file is the trained model itself. It is what you fine-tune from, and what a camera
  file is built from — but a camera cannot run it directly.
- A **`.tflite`** file is the converted copy a camera runs. This is what **Push to Device** sends.

Training normally produces both at once.

## Running on this camera

If a camera is selected in the device picker, a short list above the table shows what is actually
installed on it right now — one line per slot, with a green tick when the model is there and matches
a model you have in Zira.

Click a model's name to see the full detail: the file on the camera, where it came from, and when it
was installed.

This is reported by the camera itself, so it is the truth about that device rather than a record of
what somebody meant to send.

## What you can do with a model

Open the **⋮** menu at the end of a row.

| Action | What it does |
| --- | --- |
| Download | Saves the file to your computer. Available to administrators. |
| View training analysis | Opens the run that produced this model. |
| Push to Device | Sends the model to the selected camera, into a slot you choose. `.tflite` only. |
| Fine-tune from this model | Starts a new training run that carries on from this model instead of starting fresh. `.pt` only. |
| Build a file for the camera | Converts a `.pt` into the smaller file a camera can run. |
| Change what it looks for | Sets the model's type, which is how this list groups and filters it. Safe to correct at any time — it does not change the model. |
| Delete | Removes the model from this data source. |

## Push to Device

Use this when you want a camera to start using a model.

1. Select the camera in the device picker at the top of the page.
2. Open the **⋮** menu on the model and choose **Push to Device**.
3. Choose the slot it should go into.
4. Send it, and watch the progress in the dialog.

If the camera is asleep or briefly offline the model still reaches it — the instruction waits for up
to five minutes. The dialog confirms the install from the camera's own record, so a success here
means the camera really has the file.

If nothing is selected in the device picker, Push to Device is unavailable: there is no camera to
send to.

## Build a file for the camera

Cameras cannot run a trained `.pt` model directly, so it has to be converted first. Training usually
does this for you; use this action when you need to redo it, or want a different size-and-accuracy
trade-off.

- **Size and accuracy** — smaller files run faster on the camera; larger ones are slightly more
  accurate.
- **Calibration** — the smaller options measure themselves against a sample of your real images so
  that shrinking the model costs as little accuracy as possible. The defaults are fine.

If a warning tells you the format does not run on the camera app yet, the file will still be built
and stored — you just will not be able to push it to a device until the app supports it.

## Read next

- [Training](/docs/ai-vision/training)
- [Datasets](/docs/ai-vision/datasets)
- [Validation](/docs/ai-vision/validation)