# Model Zoo and Licences

Source: https://www.potatoannotator.com/docs/vision-spatial/model-zoo

**Potato runs several vision models. None ships with the package and none is fetched behind your back.** `potato/model_zoo.py` is one registry behind all of them, and each entry carries its task, its licence, whether that licence needs explicit acceptance, and the parameters the browser needs.

```bash
potato download-models --list
potato download-models mobile_sam
```

## What is available

| Model | Job | Runs | Size | Licence |
|---|---|---|---|---|
| `mobile_sam` | Click or box to mask | Browser | 45 MB | MIT export, Apache-2.0 upstream |
| `grounding_dino_tiny` | Text to boxes | Browser | 145 MB | Apache-2.0 |
| `sam2_video_tiny` | Prompt one frame, track the rest | Server | 181 MB | Apache-2.0 |
| `edge_sam` | Click to mask, fastest on weak hardware | Browser | — | **Non-commercial only** |
| `sam3` | Text to boxes and masks in one model | Server | ~3.5 GB | Meta SAM License |
| `edgetam` | On-device tracking | — | — | Apache-2.0 |
| `onnxruntime` | The inference runtime the others need | Browser | 13.5 MB | MIT |

Two entries have no download URL. `edge_sam` permits non-commercial use only, so making it one command away would invite installing it without reading the licence. `edgetam` has no published ONNX export at all, since upstream ships CoreML, so there is nothing to fetch yet.

## Licence gating is enforced, not documented

`download-models --list` prints `<-- NON-COMMERCIAL` next to any model whose licence needs it, and `download-models` **refuses to fetch one without `--accept-licence`**. No weights for those models ship in the package.

For a research audience this is the point worth making: a tool that tells you a model is non-commercial *before* you build a dataset with it is doing something the industry mostly does not. Discovering the licence after collection is discovering it too late.

## Air-gapped installs

Run the downloads on a connected machine and copy the model directory across. The application itself needs no network — every stylesheet, script, font and icon serves from the install, verified at **62 requests, zero external** on a rendered annotation page.

The accurate phrasing is "air-gapped once the weights are on the machine". The application needs no network; the models are a one-time transfer; and any hosted AI endpoint you configure obviously needs a route out. See [air-gapped deployment](/docs/measurement/air-gap).

## Related

- [Interactive segmentation](/docs/vision-spatial/segmentation) — `mobile_sam`
- [Text prompting](/docs/vision-spatial/text-prompting) — `grounding_dino_tiny`
- [Video tracking](/docs/vision-spatial/video-tracking) — `sam2_video_tiny`
- [Source documentation](https://github.com/davidjurgens/potato/blob/main/docs/ai-intelligence/model_zoo.md)