Models

pi can use all three. Other agents use their own models from their own services, and TurnPanel does not route them. A job runs on the one model you chose. TurnPanel does not switch it to another model on its own, except when the model's machine is offline (see The model picker).

TurnPanel includes an engine for running open models. It works on a Mac with Apple silicon and at least 5 GB of free memory, or on a computer with a graphics card with at least 6 GB of video memory. The setup offers one model sized to your machine when yours qualifies.

Find more in Settings under Models. The Library searches the catalogue and shows how each model fits your machine. Downloads resume if they are interrupted, are checked against the repository's SHA-256, and can be cancelled. You can also import a model file you already have.

Choosing a local model in the picker starts loading it at once. TurnPanel will never load a model that will not fit. It tells you the numbers instead of slowing the machine to a crawl. When memory is needed, the model used longest ago is unloaded first. Pin a model with Keep loaded to stop that.

A workspace admin adds a provider's key in Settings under Models, then Online, for the workspace or one space. The key is stored as a secret that TurnPanel adds to requests itself, so it never enters a space.

With Pro, connect your TurnPanel account in Settings under TurnPanel account. Hosted models then appear in the picker. Which spaces may use them, spending limits and usage records are handled by the TurnPanel account service (part of Pro), not by the core app.

Hosted models are part of TurnPanel's official builds. They are unavailable in air-gapped mode, and in a build from source the picker says so.

The same picker appears wherever you choose a model: new jobs, space defaults, automations and apps. It has two groups. Local lists each model with its machine and state: loaded, loading with progress, on disk, or offline. Online lists your providers' models and any hosted ones you may use. Filter by what a model can do (Sees images, Uses tools, Reasons) and by how well it fits.

If a model's machine is offline, TurnPanel tries the same model on another machine, then the space's fallback model, and otherwise tells you which machine is down.