Edit

AI Manager resource reference (preview)

This article lists ARM resource paths, properties, and defaults for AI Manager, namespace, and model deployment resources. It also lists provisioning states, runtime status fields, and the az aimanager namespace list-accesskeys output. These details apply to the AI Manager public-preview guides. The API version represented by the current generated types is 2026-09-02-preview.

Important

AKS preview features are available on a self-service, opt-in basis. Previews are provided "as is" and "as available," and they're excluded from the service-level agreements and limited warranty. AKS previews are partially covered by customer support on a best-effort basis. As such, these features aren't meant for production use. For more information, see the following support articles:

For procedures, see Deploy a model with AI Manager, Manage model deployments, and Choose a model and estimate deployment cost. To inspect command options in your installed extension, run az aimanager --help or az aimanager namespace modeldeployment --help.

AI Manager resource relationships

Resource ARM resource path Scope
AI Manager Microsoft.ContainerService/aiManagers/{aiManagerName} Resource group
Namespace Microsoft.ContainerService/aiManagers/{aiManagerName}/namespaces/{namespaceName} Child of an AI Manager
Model deployment Microsoft.ContainerService/aiManagers/{aiManagerName}/namespaces/{namespaceName}/modeldeployments/{deploymentName} Child of a namespace
AI model Microsoft.ContainerService/locations/{location}/aiModels/{modelName} Subscription and location

An AI model (catalog model) isn't nested under an AI Manager or resource group. The readable modelId identifies a catalog model. Its ARM resource name, the name value, is different.

AI Manager properties

Property Description and values Default or mutability
location Required Azure location for the AI Manager. Required when creating the resource.
properties.deletePolicy Controls what happens to the underlying cluster resources when you delete the AI Manager. Keep retains those resources; Delete deletes both the underlying cluster resources and the AI Manager resource together. Defaults to Delete when omitted.
properties.provisioningState Read-only state of the last AI Manager operation. Read-only. See Provisioning states.
properties.managedResourceGroupName Read-only name of the managed resource group created by AI Manager to hold underlying infrastructure resources. Read-only.

AI Manager namespace properties

Property Description
properties.labels Optional map of labels applied to the Kubernetes namespace.
properties.annotations Optional map of annotations applied to the Kubernetes namespace.
properties.provisioningState Read-only state of the last namespace operation. See Provisioning states.

AI Manager model deployment properties

Property Description and values Update behavior
properties.modelResourceId Full ARM resource ID of the catalog model used by the deployment. Immutable after creation. Create a new deployment to use a different catalog model.
properties.vmSize VM SKU used by the deployment. Choose a feasible plan from az aimanager model calculate-cost. Immutable after creation. Create a new deployment to use a different VM size.
properties.scale.manual.replicas Desired fixed replica count. Change with az aimanager namespace modeldeployment update --replicas.
properties.scale.autoscale.minReplicas and properties.scale.autoscale.maxReplicas Minimum and maximum replica counts for autoscaling. Change with --min-replicas and --max-replicas. Don't combine these options with --replicas.
properties.performanceMode Performance mode. Values are Balanced (default balance between latency and throughput), Latency (optimize for low request latency), and Throughput (optimize for aggregate throughput). The CLI update command exposes --performance-mode.
properties.status Read-only runtime status. Read-only. See Runtime status.

The CLI update command exposes scaling and performance mode options. To change a deployment's catalog model or VM size, create a new model deployment.

Select an AI Manager model from the catalog

Before using the catalog, register the AI Manager preview feature. Run az aimanager model list for the subscription and location where you plan to deploy. Choose a catalog model by its readable modelId. Use its matching name value as the ARM resource name for the cost calculator and when building the deployment's modelResourceId.

az aimanager model list \
  --subscription "<subscription-id>" \
  --location "<location>" \
  --output table

For model selection and cost planning, see Choose a model and estimate deployment cost.

Provisioning states

The AI Manager, namespace, and model deployment resources use these provisioning state values:

State Description
Creating The resource is being created.
Updating The resource is being updated.
Deleting The resource is being deleted.
Succeeded The last operation completed successfully.
Failed The last operation failed.
Canceled The last operation was canceled.

Runtime status

properties.status is read-only. Each field in the following table is under that path, for example properties.status.currentReplicas. Current model deployment API types include these fields:

Field Description
endpoint Serving endpoint of the model deployment, reported by the runtime. Full path: properties.status.endpoint.
currentReplicas Number of ready replicas currently serving traffic.
desiredReplicas Desired replica count reported by the controller. For manual scaling, this matches properties.scale.manual.replicas; for autoscaling, it is the current target derived by the autoscaler.
engine and engineVersion Inference engine name and version.
maxModelLen Maximum model context length, in tokens, configured for the deployment.
quantization Quantization level applied to the model weights.
peakTokensPerMinute Peak tokens per minute measured by a live stress test.
estimatedProvisionTimeSeconds Estimated total time, in seconds, for the deployment to become ready, including VM provisioning, image and model-weight pulls, and engine warm-up.

Separately from properties.status, the az aimanager namespace list-accesskeys command returns the following namespace access information:

Field Description
endpoint In the list-accesskeys output, the OpenAI-compatible base URL for the namespace inference endpoint. Use it as returned by the command.
primaryKey and secondaryKey API keys for the endpoint. Treat both values as secrets.
lastRotatedAt UTC timestamp of the last key rotation. Absent until a rotation occurs.

Next steps