Skip to main content
Model Nodes are managed clusters that run participant models and make them callable over the network — without exposing the underlying model code. They solve three problems:
  • Remote access — models for a Crunch are callable via gRPC from anywhere
  • Intellectual property protection — models are only reachable through an API; no direct code access is provided
  • Access control — the protocol defines who can call which model, and under what rules

​
How it works

Here is what happens when a Cruncher submits a model:
1

Model submission

A Cruncher submits their model code to a Crunch run by a Coordinator.
2

On-chain authorization

The Crunch Protocol writes an authorization on-chain, granting the right to run and call that model remotely.
3

Deployment

The Model Orchestrator reads the blockchain, detects the new authorization, and reacts by:
  • Building an execution environment (e.g., a Docker container) that makes the model callable
  • Deploying the container on a cloud platform
  • Sharing connection details with the Coordinator
4

Inference

The Coordinator uses the Model Runner Client (a Python library) to connect to the deployed model and:
  • Fan out requests to many models concurrently
  • Feed data into models
  • Collect predictions

​
Components

Model Nodes rely on three open-source components maintained by CrunchDAO:
ComponentRoleLinks
Model RunnergRPC server that runs alongside each model, making it callable over the networkDocs · GitHub
Model Runner ClientCoordinator-side Python library for calling models concurrentlyDocs · GitHub
Model OrchestratorManages the lifecycle of model containers — starting, stopping, and health monitoringDocs · GitHub

​
Deep dives

Model Runner and Client

How models are made callable and how the Coordinator calls them.

Model Orchestrator

How model containers are deployed, monitored, and kept reachable.

Access Control

The Secure Model Protocol that authenticates communication.