# Model Discovery

Find a model, install it on the computer that will run it, then load it into memory. A registry listing, an installed file, and a loaded model are different states.

## Choose a model and variant

Browse [Models](/models) or the [Actual model registry](https://models.actual.inc). The [machine-readable index](https://models.actual.inc/index.json) links to the current models, revisions, variant manifests, and downloadable artifacts.

Use each model's published variants as the source of truth. Available quantizations vary; do not assume every model has a Q6 variant. Q8_0 generally uses more disk and memory than Q4_K_M. Speed depends on the model and hardware, so a smaller download is not a promise of faster inference.

Download size is not the total memory needed to run a model. Context length, runtime allocations, and other workloads also consume memory.

## Download from the website

On a model page, choose **run model**. In **Download to cluster**, select an eligible online cluster, optionally enable **Load after install**, and choose **Queue install**.

Watch the displayed progress and any failure message. A queued download has not finished installation. If no eligible cluster appears, check that your computer is online and that you have the necessary ownership or organization permissions. See [Organizations](/docs/organizations).

## Download and load from the CLI

Find the desired model:

```bash
actual models search "gemma"
```

Registry identifiers use this format:

```text
actual:publisher/model@revision/QUANT
```

Copy the current complete identifier and quantization from the registry. For example, this pins a specific published revision:

```bash
actual models download actual:google/gemma-4-12b-it@5926caa4ec0cac5cbfadaf4077420520de1d5205/Q4_K_M
```

Wait for the download to finish, then inspect and load it:

```bash
actual models jobs
actual models list
actual models load
actual status
```

The interactive load picker lists installed models. In a script, pass the stable selector shown by `actual models list` as the argument to `actual models load`; do not rely on a display name that could identify multiple variants.

The download command also accepts supported Hugging Face repository identifiers. Models outside the Actual registry may need additional compatibility checks or files.

If zsh shows `dquote>`, the shell is waiting for a closing quote and has not run the command. Press **Ctrl-C**, then run `actual models load` to use the picker.

## Progress, cancellation, and unloading

Downloads belong to the daemon and can continue after you close the terminal. Use `actual models jobs` to inspect them and `actual models cancel JOB_ID` to cancel a job, replacing `JOB_ID` with its listed ID.

Job history can be unavailable after a restart. Use `actual models list` to check installed files after restarting. Rerun a failed download to retry it.

Use `actual models unload` to free the loaded model's memory. Unloading is separate from deleting installed files and can interrupt inference.

## Available and loaded inventory

Only online devices contribute to the public API's current inventory. `GET /v1/models` returns the models advertised by those devices, including models available to load and models already loaded.

Without `X-Cluster-ID`, inventory is aggregated across your account's clusters. An entry's `loaded` value means it is loaded somewhere in that result; inspect `clusters[].cluster_id` and `clusters[].loaded` before pinning a request to a particular cluster. A model loaded in one cluster is not automatically loaded in another.

Use the returned `id` or stable `selector` for inference. Display names can become ambiguous when several installed variants share a name.

The public inference API does not download or load models. Use the CLI or the website, then query inventory again and wait for the desired cluster to report the model loaded. See [API Quickstart](/docs/api-quickstart).

## Load from Chat

In [Console > Chat](/console/chat), choose the cluster, then an installed model. The picker reflects that cluster's online devices and marks loading and loaded state. Selecting an unloaded model requests a load; wait until live state confirms readiness before sending.

If a device disconnects, inventory may disappear or change. Restore the connection before treating a saved model selection as ready.

## Vision and audio

Check the loaded model's capabilities for your client and hardware. Gemma 4 E2B IT is a registry option for image and audio use; download it through Actual so its companion files are included. Other variants may support images without audio.

On macOS, check the eye and waveform icons after loading. Console Chat accepts text only. See [Vision & Audio](/docs/vision-and-audio) for compatible applications and attachment guidance.