Local AI
The Local Model Center is where you manage local AI on this machine. You can browse, install, import, remove, and inspect Runtime's local assets — but installing something here never picks which implementation answers an App's request.
Surface Boundary
| Concept | Meaning |
|---|---|
| Local asset catalog | Runtime-issued inventory of installable assets |
| Installed assets | Resources registered on this machine |
| Installation progress | Runtime-issued download, verification, and materialization state |
| Recommendation feed | Runtime-ordered installation suggestions and evidence |
| Capability intent | An owner's Local or Cloud preference in AIConfig |
The asset catalog is machine setup, nothing more. Picking a bundle in the Local Model Center means "install this" or "remove this" — it doesn't pin a model, engine, or route for later App calls.
Runtime Truth Projection
Everything on this page comes from Runtime; Desktop doesn't reconstruct or second-guess it locally.
| Concern | Owner |
|---|---|
| Catalog and installed-asset inventory | Runtime |
| Download, verification, and materialization status | Runtime |
| Recommendation order and evidence | Runtime |
| Device and dependency diagnostics | Runtime |
| Implementation selection for each request | Runtime |
Recommendations appear in the order Runtime gives them, with Runtime's own reasons and compatibility evidence attached. Desktop doesn't score, grade, group, or reshuffle the list on the client.
Dependency Setup
Engines that need system dependencies install through a Runtime materializer. You confirm, and Runtime runs the whole operation — download, verification, installation, and cleanup. Desktop shows typed progress and failures along the way; it never executes arbitrary PowerShell or shell commands itself.
A finished dependency install is machine-setup evidence. It isn't a signal that a particular model is ready, and Apps can't use it to pick an implementation.
Reader Scenario: Install a Local Asset
- Open Local Model Center. Desktop reads the Runtime catalog and installed inventory.
- Browse. The user reviews Runtime-issued metadata and recommendation evidence.
- Install. The user selects an asset bundle for installation. Runtime downloads, verifies, and registers it.
- Inspect result. Desktop projects the typed installation result or failure.
- Use a capability. An App with Local capability intent submits its ordinary request without a model, route, connector, or target. Runtime decides whether an admitted installed asset can service it.
Installation does not create an App-visible binding and does not guarantee that a particular implementation will handle the next request.
Reader Scenario: Local Memory Capability
An owner may express Local intent for an admitted memory or embedding capability. The configuration contains the owner identity, capability, and Local intent only. Runtime chooses the embedding implementation and owns any bank migration or cutover required by machine configuration. Desktop does not expose an embedding-model picker or a client-computed readiness state.
Realm Connectivity
Your Realm connection and your local Runtime are separate things. A Realm outage doesn't by itself disable local AI. Local execution still needs the Runtime to be reachable, your Local preference to allow it, and Runtime to find a valid implementation. When a prerequisite is missing you get a typed failure — never a fabricated success.
Key Points
- The Local Model Center manages machine assets owned by Runtime.
AIConfigrecords your Local or Cloud preference per capability.- Apps get no model activation, warming, engine binding, route readiness, or per-model health controls.
- Runtime alone picks the implementation for each request.
- Runtime diagnostics and recommendation evidence never become request input.