ML Models

What are ML Models?

ML Models in Cyberwave are AI models registered in your workspace that can process various inputs — video, images, audio, text, or robot actions. They integrate with workflows and can run in the cloud or on edge devices.

ML Models define what the model can do (input types) and where it runs (provider). The actual inference happens through the model provider’s API or on your edge device.

Model Capabilities

Each ML Model specifies what inputs it can process:

Capability	Description	Example Use Cases
`can_take_video_as_input`	Process video streams	Surveillance, teleoperation
`can_take_image_as_input`	Process single images	Quality inspection, object detection
`can_take_audio_as_input`	Process audio data	Voice commands, anomaly detection
`can_take_text_as_input`	Process text prompts	Natural language commands
`can_take_action_as_input`	Process robot actions	Behavior cloning, RL policies

Model Providers

Models can run through different providers:

Local / Edge

Run on your edge devices using ONNX, TensorRT, or custom inference

Cloud APIs

Use OpenAI, Anthropic, or other cloud AI services

Hugging Face

Deploy models from Hugging Face Hub

Custom

Your own inference servers and endpoints

Registering a Model

Dashboard
Python SDK

Navigate

Go to ML Models in your workspace.

Add model

Click Add Model.

Configure

Fill in the model details: name, description, external ID (model identifier for the provider), provider name (e.g. openai, local, huggingface), and input capabilities.

Create

Click Create.

from cyberwave import Cyberwave

cw = Cyberwave(api_key="your_api_key")

model = cw.api.create_mlmodel({
    "name": "Object Detector",
    "description": "YOLOv8 model for detecting warehouse items",
    "model_external_id": "yolov8n",
    "model_provider_name": "local",
    "can_take_image_as_input": True,
    "can_take_video_as_input": True,
    "visibility": "private",
    "tags": ["vision", "detection", "warehouse"]
})

print(f"Model UUID: {model.uuid}")

Model Visibility

Visibility	Who Can Access
`private`	Only your workspace members
`workspace`	All workspace members
`public`	Anyone (admin-only to create)

Using Models in Workflows

ML Models integrate with workflow nodes for automated processing:

Try it in the Playground

Every model detail page (/{workspace-slug}/models/{model-slug} or /models/{uuid} for models without a slug) has an interactive Playground tab. Gemini Robotics-ER renders detected points as an overlay on your image, VLMs stream back text, im2mesh models preview the generated GLB inline, and edge/VLA models surface the exact CLI + SDK commands needed to run them locally. See Model Playground.

Running Inference

Cloud Models
Edge Models

For cloud-based models, Cyberwave routes requests to the provider:

response = cw.api.vlm_generation({
    "model_uuid": model.uuid,
    "prompt": "What objects do you see? How should the robot pick them up?",
    "image_url": "https://..."
})

Listing Models

from cyberwave import Cyberwave

cw = Cyberwave(api_key="your_api_key")

models = cw.api.list_mlmodels()

for model in models:
    print(f"{model.name} ({model.model_provider_name})")
    print(f"  Video: {model.can_take_video_as_input}")
    print(f"  Image: {model.can_take_image_as_input}")
    print(f"  Text: {model.can_take_text_as_input}")

Vision-Language-Action (VLA) Models

VLA models combine vision, language understanding, and action generation for end-to-end robot control. Cyberwave provides infrastructure for running VLA model inference and training on Cloud Nodes.

Supported VLA Models

Model	Description	Deployment
SmolVLA	Lightweight VLA from HuggingFace LeRobot	Cloud Node
OpenVLA	Open-source VLA model	Cloud Node

Model Weights

VLA models store their weights in Cyberwave and expose them via the MLModel API:

GET /api/v1/mlmodels/{uuid}/weights
→ { "signed_url": "...", "expires_at": "..." }

This allows Cloud Nodes to fetch model weights on demand and cache them locally.

Running VLA Inference

VLA inference runs on Cloud Nodes using the CwProcessor orchestrator:

Weights Download - Fetched from MLModel API via signed URLs
Camera Binding - Background threads continuously fetch camera frames
Control Loop - Observe → Predict → Execute cycle
Action Publishing - Predicted actions sent to robot via MQTT

VLA Cloud Node Guide

Learn how to build and deploy VLA models on Cyberwave Cloud Nodes

Concepts

Platform Features

Cyberwave Edge

Technical Reference

Use-Case Recipes

What are ML Models?

Model Capabilities

Model Providers

Local / Edge

Cloud APIs

Hugging Face

Custom

Registering a Model

Model Visibility

Using Models in Workflows

Try it in the Playground

Running Inference

Listing Models

Vision-Language-Action (VLA) Models

Supported VLA Models

Model Weights

Running VLA Inference

VLA Cloud Node Guide

Concepts

Platform Features

Cyberwave Edge

Technical Reference

Use-Case Recipes

Documentation Index

​What are ML Models?

​Model Capabilities

​Model Providers

Local / Edge

Cloud APIs

Hugging Face

Custom

​Registering a Model

​Model Visibility

​Using Models in Workflows

​Try it in the Playground

​Running Inference

​Listing Models

​Vision-Language-Action (VLA) Models

​Supported VLA Models

​Model Weights

​Running VLA Inference

VLA Cloud Node Guide

What are ML Models?

Model Capabilities

Model Providers

Registering a Model

Model Visibility

Using Models in Workflows

Try it in the Playground

Running Inference

Listing Models

Vision-Language-Action (VLA) Models

Supported VLA Models

Model Weights

Running VLA Inference