Dynamic AI Model

Overview

The Dynamic AI Model node enables runtime switching between multiple machine learning models within a single flow. Unlike the standard AI Model node, which uses a single preloaded model, this node can unload the current model and load a new one on demand — useful when your workflow must handle varied tasks within the same deployment or when GPU memory cannot support multiple models simultaneously.

Use Dynamic AI Model when:

• You need to switch between different inspection tasks based on external triggers.
• You want to process multiple object types that require specialized models without running them all simultaneously.
• Your hardware resources (GPU memory) are limited and cannot support multiple concurrent model instances.

💡 Tip

At deployment startup, no model is loaded by default. The node must receive a valid model name on its model_name input before it begins processing images. Switching between models takes a brief moment during which no images are processed.

Inputs

Name	Type	Description
image	image	The source image to be processed by the active model.
model_name	string	The identifier of the model to activate. Format: CustomerName/ModelName:Version (must match a pre-selected model).

Copy the model name from the Models page — it follows the CustomerName/ModelName:Version format.

Outputs

The available outputs depend on the Model output format parameter.

Mapped format — all results bundled into a single output:

Name	Type	Description
prediction	prediction	A dictionary of prediction results from the active model, organized by prediction class.

Unpacked format — individual attributes as separate outputs, prefixed by the prediction class name (e.g. Defect_polygons, Defect_probability):

Name	Type	Description
{Class}_heatmap	heatmap	Spatial distribution of detected features.
{Class}_points	list[point]	Detected coordinates from the active model.
{Class}_polygons	list[polygon]	Geometric outlines of objects detected by the active model.
{Class}_class	string	The label of the detected prediction class.
{Class}_probability	number	Confidence score of the classification (0–1).

Always present:

Name	Type	Description
current_model	string	The name of the currently active model, or an error message if none is loaded.

💡 Tip

The specific unpacked outputs (heatmap, points, polygons, class, probability) depend on what the selected model produces.

Parameters

Parameter	Default	Valid values	Tunable
Select AI models	[]	Multiselect	❌
Model output format	unpacked	unpacked, mapping	❌

Select AI models

A list of pre-selected models that the node is authorized to load at runtime. Only models chosen here can be activated via the model_name input.

⚠️ Important constraints

• All selected models must have the same inputs and outputs (output classes, types and constraints).
• Models must be downloaded and marked as valid in the Models section to appear as options.
• Any selected model must not be used in any other AI Model or Dynamic AI Model node.

Model output format

Determines how prediction data is exposed to downstream nodes.

• All value types (unpacked) — individual attributes (polygons, probability, class, etc.) appear as separate output pins.
• Only model classes (mapping) — all results are bundled into a single prediction output, organized by class. Use Extract Attributes to access individual values.

Example

Multi-product inspection line

A packing line processes both apples and oranges. An external sensor identifies the product type and sends the corresponding model name to the node.

Configuration:

• Select AI models: Apple_Detector:1, Orange_Detector:1
• Model output format: unpacked

Result: When the node receives Apple_Detector:1, it unloads any previous model, loads the apple inspection model, and outputs polygons outlining bruises on the fruit.

Displaying the active model on the dashboard

The current_model output shows the name of the currently active model. Connect it to a Text output node with a Value widget to display it on the dashboard.

Displaying the active model name on the dashboard via the Text output node.

Normal operation — the dashboard displays the active model name.

'No model registered' — no model has been received yet or the name didn't match.

"No model registered" appears when:

• The node is in its default state (e.g. after deployment startup or configuration reset) and has not yet received a model name.
• The received model_name doesn't match any of the selected models (e.g. contains a typo). The previously active model is unregistered and the node waits for the next valid input.

💡 Tip

If "No model registered" appears unexpectedly, verify that the strings sent into the model_name input exactly match the names of the selected models. The message may also appear briefly during system initialization — this is normal.

Switching models via REST API

You can send the model name via an external REST API request.

Prerequisites:

• An external Connection to the server.
• A String Input node mapped to the connection in Deployments.
• The String Input node must be set as Optional in Deployments.

Set the String Input as optional in Deployments.

Example flow configuration for model switching via external REST API.

Example request:

curl -k -X POST \
  'https://SERVER_URL/hashi/v1/http/external/input/PROJECT_ID/CONNECTION_NAME' \
  -H 'X-MQ-Payload-Type: string' \
  -H 'Content-Type: application/json' \
  -H 'X-Correlation-ID: dce5d400-7513-4afb-bb90-716b2cc86fd7' \
  -d '{"value": "CustomerName/ModelName:Version"}'

Related links

Reference:

• AI Model — run inference on a single, fixed model
• Extract Attributes — extract specific attributes from model predictions
• Polygon Threshold — filter detected objects by size or confidence
• Models — manage model downloads and availability
• Connections — set up external connections for API-driven workflows
• Deployments — configure deployment input mappings