Create LLM Model

Use the Create LLM Model form to add a physical provider model to the LLM global Model Catalog shared by every host. Open it from Marketplace > LLM Model Catalog and choose Create. Only platform catalog administrators should have permission to use this command.

This guide applies only to /app/form/createLlmModel. For registrations, deployments, aliases, policies, and publication, see the LLM Model Control Plane guide.

Required Fields

Model Version is optional. active is backend-managed: create and update keep the model active, while the delete command soft-deletes it.

Structured Fields

Modalities, Operations, and Declared Capabilities are stored as typed arrays or objects. Do not enter JSON or YAML as a quoted string.

Modalities and Operations

These fields open on the Form tab. Add one string per array item. You can also use the JSON or YAML tab, for example:

["text", "image"]

- generate
- embed

Declared Capabilities

This open-ended object starts on the JSON tab. The Form tab is unavailable because the current schema does not prescribe capability property names. Enter an object whose keys and values describe the provider model, for example:

{
  "streaming": true,
  "tools": true
}

The equivalent YAML is:

streaming: true
tools: true

After changing JSON or YAML, choose Apply. Apply parses the draft, checks its root type and schema constraints, and updates the form model only when it is valid. Choose Reset to discard the draft and restore the last valid value. The Create action remains blocked while a structured draft is invalid or has not been applied.

Categories and Tags

Categories and tags are optional. The selectors show active global taxonomy values registered for the llm_model entity type. Host-specific taxonomy cannot be assigned to a global model.

NVIDIA Nemotron embedding example

For the light-knowledge demo, use these values for the NVIDIA catalog Model:

Use this Declared Capabilities object:

{
  "operations": ["embed"],
  "embedding": {
    "maxBatchItems": 1,
    "maxInputTokensPerItem": 4096,
    "maxAggregateInputTokens": 4096,
    "supportedDimensions": [2048],
    "supportedEncodings": ["float"],
    "maxResponseBytes": 16777216,
    "space": {
      "spaceId": "nvidia-nemotron-3-embed-1b-float-v1",
      "revision": 1,
      "dimension": 2048,
      "normalization": "l2",
      "distanceMetric": "cosine",
      "documentInputTransformVersion": "document-v1"
    }
  }
}

NVIDIA currently documents only the native 2048-dimensional output for this hosted model. Document indexing must use passage semantics and retrieval must use query semantics. The catalog capability declaration records that contract; the gateway provider adapter must still send the correct provider request. See the NVIDIA NIM support matrix.

Create the Record

Review the values and choose Create LLM Model. The form sends the createLlmModel command and preserves modalities and operations as arrays and declaredCapabilities as an object. After a successful command, the browser returns to GenAI Admin > LLM Models.

Common Problems

• Create is blocked after editing JSON or YAML: choose Apply to commit the draft, or Reset to discard it.
• JSON or YAML error: correct the highlighted syntax and choose Apply again. The last valid value remains unchanged.
• Required-field validation: provide every required field and use positive integers for both token limits.
• No categories or tags are available: confirm that global taxonomy values are active and registered for llm_model.
• 403 on Create: confirm access to the lightapi.net/genai/createLlmModel/0.1.0 command endpoint with the required write scope and role permission.