Register a model

Registering creates the model record on Reactor. You do it once per model; releases (image and weights pairs) are published against it afterwards. Models are private to the Reactor account that creates them and its API keys. Contact Reactor if you want to make a model publicly available. You do not need to run register before a normal workspace deployment. reactor model deploy creates the record when needed, then publishes and deploys the release. Use the register command when you only want to reserve the model name or inspect this step separately.

Register from your workspace

Run it from inside your workspace:

cd my-model
reactor model register

The CLI walks up from the current directory until it finds your reactor.yaml (stopping at the repository root) and reads your model details from it, so no flags are needed. Those details are the source of truth for everything Reactor knows about your model:

reactor.yaml

model:
  # Model name: `<org>/<name>`, lowercase. A short slug without the
  # org prefix also works; Reactor fills in your org.
  name: my-awesome-model

  # Release tag used by publish and deploy. Full semver, e.g. v1.0.0.
  version: v0.0.1

  description: |
    A vision transformer for partner integration.

  # Optional accelerator and CPU/memory overrides.
  # See "Resources and accelerators" below.
  resources:
    gpu:
      type: NVIDIA_H100
      count: 1

runtime:
  import: model:MyModel
  config: config.yaml

The success output prints the canonical model name and points at the next command:

Registering model 'my-awesome-model'...
✓ Model registered successfully
  ID:     a1b2c3d4-...
  Name:   vendor/my-awesome-model
  Status: ACTIVE
  Public: false

Next step - deploy the workspace:
  reactor model deploy

The Name in the output, vendor/my-awesome-model, is your model’s full identifier on Reactor. It is the value commands take in a <model>:<release> argument. Inside a workspace you rarely type it: the commands read it from reactor.yaml for you.

Register without a workspace

For one-off invocations, pass the model name on the command line:

reactor model register --name vendor/my-awesome-model

A short slug without the org prefix (--name my-awesome-model) also works; Reactor fills in your org.

Resources and accelerators

The accelerator your model runs on, and any CPU or memory overrides, all live under model.resources in reactor.yaml:

reactor.yaml

model:
  resources:
    # Accelerator. Omit for CPU-only models.
    gpu:
      type: NVIDIA_H100
      count: 1

    # Optional CPU/memory overrides. Omit to use Reactor defaults.
    cpu:
      request: 750m
      limit: "2"
    memory:
      request: 8Gi
      limit: 32Gi

CPU and memory quantities follow Kubernetes conventions:

Field	Example	Meaning
`cpu.request`	`500m`	Half a CPU requested
`cpu.limit`	`2`	Up to two CPUs
`memory.request`	`4Gi`	4 GiB requested
`memory.limit`	`32Gi`	Up to 32 GiB memory limit

The accelerator types Reactor accepts as gpu.type:

Family	Types
NVIDIA GPU	`NVIDIA_A100`, `NVIDIA_H100`, `NVIDIA_H200`, `NVIDIA_B200`, `NVIDIA_GB200`, `NVIDIA_B300`, `NVIDIA_RTX6000`, `NVIDIA_RTX6000_PRO`
AMD GPU	`AMD_MI300X`, `AMD_MI350X`, `AMD_MI400`
AWS Trainium	`AWS_TRAINIUM_2`, `AWS_TRAINIUM_3`
AWS Inferentia	`AWS_INF_2`

The platform is authoritative for this list. Print the currently accepted values with reactor gpu ls (pass --refresh to bypass the local cache after a platform release adds a new accelerator). To change these settings on an existing model, edit reactor.yaml and run reactor model update.

Deploy a model

Publish the workspace and activate its release.

Manage the model

Update the spec, list running instances, or delete the model.

Get Started

CLI & Platform

Build a Model

Media & Transport

Releases

Register a model

Register from your workspace

Register without a workspace

Resources and accelerators

Next

Deploy a model

Manage the model

​Register from your workspace

​Register without a workspace

​Resources and accelerators

​Next

Deploy a model

Manage the model

Register from your workspace

Register without a workspace

Resources and accelerators

Next