The model layer

Models are cards that carry trained weights instead of a fixed calculation.

They are cards

A model has ports, parameters and a manifest like any other card. The difference is that its behaviour comes from a file of weights, which is versioned and distributed with it.

Why they are separate in the marketplace

Weights are large, they cannot be read the way a Python file can, and their provenance matters more. Listing them as their own type makes that visible rather than burying a 200 MB opaque file among the filters.

What to check before using one

• What it was trained on, and whether that resembles your recordings.
• Whether the author published the training procedure.
• What it does when given something outside its training distribution. Most models answer confidently rather than refusing.