4 of 7 · Serving models

Deployments and predictions

Create a deployment from an eligible model version, call its predictor, and inspect usage and result media.

Move from model version to prediction#

Select an eligible result
Use a completed model version with a deployable checkpoint. Record both the pipeline and model-version identifiers.
Create or retrieve the deployment
Avoid duplicate deployments for the same intended model version. Confirm active state before routing production traffic.
Submit representative input
Use the predictor contract for the deployed task and media type. Keep large files and protected inputs server-side.
Poll asynchronous predictions
When the predictor returns accepted work, retain the usage or task identifier and poll its documented status endpoint.
Retrieve protected artifacts
Load previews, explainability output, or result media with authentication through the documented result paths.
Monitor and deactivate intentionally
Review usage, latency, failures, and consumers before changing active state.

Deployment capacity and task support depend on the organization plan and enabled runtime services.
Prediction response and artifact shapes vary by model family; use the endpoint-specific card.
Treat a deactivation as an externally visible change and coordinate it with API consumers.