Model resources
A model resource is a routable instance available to an API configuration. It can be a platform model provided directly by Lumoswitch or a connected model discovered from an upstream.
Platform model ───────────────────────────────────────┐
├─> model resource -> candidates -> routing
Connected upstream -> discovery -> Connected model ──┘Platform models and connected models differ only by source. Downstream clients always use an API configuration, its runtime policy, and its access key. The same model ID may appear through the platform, multiple upstreams, or multiple credentials.
Discover models
- Open Model resources and review platform models.
- If a connected upstream exists, run discovery.
- Review model ID, source, and health.
Results persist between requests. Run discovery again after provider permissions, keys, or available models change.
Resource fields
| Field | Meaning |
|---|---|
| Model ID | Provider's real model identifier |
| Display name | Human-readable catalog name |
| Provider / source | Platform model or the connected model's upstream |
| Health | Latest discovery or request signal |
| Routable | Whether configurations may select this resource |
| Capabilities / context | Metadata enriched from the model catalog |
Health states
| State | Meaning |
|---|---|
| Healthy | The latest discovery or call succeeded |
| Unhealthy | A recent discovery or call failed |
| Stale | The last health signal is too old and discovery should run again |
| No longer returned | The upstream stopped listing this model |
| Unknown | No sufficient health signal exists yet |
Catalog matching
Lumoswitch matches the inference provider and provider model ID to a provider model, then reads its canonical catalog model to add context length, capability tags, external reference pricing, and catalog status. Reference prices are display metadata and never determine the model price charged to a user. Catalog metadata helps selection but does not determine whether the provider account can call the model.
Client-facing names and upstream model IDs
Configurations select model resources, then expose stable client-facing model names:
Client-facing model name: lumos-chat
-> DeepSeek / deepseek-chat
-> OpenAI / gpt-4o-miniRelationship to API configurations
Configurations select platform model or connected model resources; they never issue a special key that bypasses routing for platform models. Remove a resource from candidate lists before deleting it. Rerun discovery after provider permissions or keys change.
When discovery returns nothing
- Confirm whether the platform currently provides an available platform model.
- For connected models, run the upstream connection test and verify list/invoke permissions.
- Check the base URL, provider quota, organization, and region restrictions. Connected models are not limited by the Lumoswitch account balance.
- Quickstart may continue without a discovered model, but calls require a healthy candidate before use.