Deploy models to a sandbox

Deploy a model from Registry to an isolated CoreWeave sandbox environment for testing and iteration. Send inference requests to the deployment endpoint with tools such as curl or a Python HTTP client.

Deploying models from Registry to sandboxes is in private preview and currently requires GPU-enabled sandboxes. To request access, contact your account team or email support@coreweave.com.

CoreWeave Sandboxes have a limited lifetime and stop when they expire. When a sandbox expires, CoreWeave releases its compute resources and its endpoint becomes unavailable. To create another sandbox, clone the deployment. Cloning creates a new sandbox with a new generated endpoint URL. Deployments are scoped to your organization and associated with the team you select. To create, clone, stop, or otherwise manage deployments, you must have a Registry role with write access in the organization. Your organization must also have access to sandboxes and GPU sandboxes, available sandbox quota, and sufficient sandbox compute.

Deploy a model

Follow the steps below to deploy a model to a sandbox:

Navigate to Registry.
Select a registry.
Select a collection.
Select a model version.
Select the action menu (), then select Deploy.
From the Version dropdown, select the model version to deploy.
Enter a name for the deployment for the Deployment name field.
Select Next.
Select the team to associate with the deployment.
Select a compute resource.
From the Lifetime dropdown, select how long the sandbox can run.
Configure the Serving engine. The default and currently supported engine is vLLM.
Select Next.
Review the deployment details, then select Deploy.

Compute resources

The compute resource determines the CPU, memory, and GPU resources allocated to the sandbox.

Type	Description
CPU (vCPU)	The number of virtual CPUs allocated to the sandbox.
Memory (GiB)	The amount of system memory allocated to the sandbox.
GPU	The number of GPUs allocated to the sandbox.

GPU sandboxes must be enabled for your organization. Until GPUs are enabled, a request that includes a GPU fails with CWSANDBOX_GPU_NOT_ALLOWED and the message this organization is not entitled to create GPU sandboxes.

Sandbox states

A sandbox can have one of the following states:

State	Description
Starting	CoreWeave is creating the sandbox and deploying the model.
Running	The sandbox is running and available for use.
Stopping	CoreWeave is stopping the sandbox.
Stopped	A user stopped the sandbox. Its endpoint is no longer available.
Expired	The sandbox reached its expiration time. Its endpoint is no longer available.
Failed	An error prevented the deployment from starting.

View a deployment

View the status, expiration time, configuration, and other details for a deployed sandbox.

Navigate to Registry.
Select a registry.
Select a collection.
Select a model version.
Select the Deployments tab.
Open the deployment details by doing one of the following:
- Select the name of the deployed sandbox.
- Select the action menu () and choose View deployment details.

Deployment overview

The deployment overview includes the following information:

Status: The current state of the sandbox.
Created by: The user who created the deployment.
Created at: When the deployment was created.

Sandbox configuration

The sandbox configuration includes the following information:

Instance name: The name of the sandbox instance.
Serving engine: The engine that serves the model.
Team: The team that created the sandbox.
Compute resources: CPU, memory, and GPU resources allocated to the sandbox.
Expires in: The remaining time before the sandbox expires.
Endpoint address: The address used to send requests to the deployed sandbox.
curl example: An example request to the deployment endpoint using curl.

Access a deployed model

Send requests to a deployed model by using the endpoint address in the sandbox configuration. The endpoint remains available until the sandbox expires or you stop the deployment.

Sandbox endpoints are publicly accessible and don’t support Forge API key authentication. Anyone with the endpoint URL can send requests to the deployed model. Don’t use a sandbox endpoint for models or data that require access controls.

Find the endpoint address

You can find the endpoint address for your deployed sandbox in the sandbox configuration section of the deployment overview.

Open the deployment details.
In Sandbox configuration, find Endpoint address.

Test the deployment

Send a chat completion request to verify that the deployment can serve inference requests.

Python
curl

Copy and paste the following Python code to test your deployed sandbox. Set base_url to the endpoint address of your deployment and install httpx if you haven’t already done so by running pip install httpx.

import httpx

base_url = "sandbox-url"

response = httpx.post(
    f"{base_url}/v1/chat/completions",
    headers={
        "Content-Type": "application/json",
    },
    json={
        "model": "local",
        "messages": [
            {
                "role": "user",
                "content": "Reply with OK."
            },
        ],
        "temperature": 0,
        "max_tokens": 5,
    },
    timeout=60,
)

response.raise_for_status()
print(response.json()["choices"][0]["message"]["content"])

Copy and paste the following curl command to send a test request to your deployed sandbox. Replace sandbox-url with the endpoint address of your deployment:

curl --fail-with-body \
  "sandbox-url/v1/chat/completions" \
  --header "Content-Type: application/json" \
  --data '{
    "model": "local",
    "messages": [
      {
        "role": "user",
        "content": "Reply with OK."
      }
    ],
    "temperature": 0,
    "max_tokens": 5
  }'

Clone a deployment

Clone a deployment to create a new sandbox with its own generated endpoint URL. By default, the cloned deployment uses the original deployment name with -clone appended. You can change this name before confirming the deployment. Deployment names must be unique among deployments that are starting or running. You can reuse a name after the existing deployment stops or expires. To clone a deployment, follow these steps:

Navigate to the Registry.
Select a registry.
Select a collection.
Select a model version.
Select the Deployments tab.
Select the action () menu, then select Clone deployment.
Review or update the deployment settings, then confirm the deployment.

Stop a deployment

Stop a running deployment. You can’t restart a stopped deployment, and its endpoint becomes unavailable. To serve the model again, create a new deployment. To stop a running sandbox deployment, follow these steps:

Navigate to Registry.
Select a registry.
Select a collection.
Select a model version.
Select the Deployments tab.
Select a deployment.
Select the action menu (), then select Stop deployment.

Billing and usage

Deploying a model to a sandbox consumes CoreWeave sandbox compute. Model workflows can also incur product-specific usage when they use other W&B services, such as Weave, Inference, or Training. Organization and billing admins can review plan details, monitor usage, configure usage and spending alerts, manage payment methods, and access invoices from Billing settings. If your account has insufficient sandbox compute, you might need to purchase additional usage, upgrade your plan, or contact CoreWeave support at support@coreweave.com.

​Deploy a model

​Compute resources

​Sandbox states

​View a deployment

​Deployment overview

​Sandbox configuration

​Access a deployed model

​Find the endpoint address

​Test the deployment

​Clone a deployment

​Stop a deployment

​Billing and usage

Deploy a model

Compute resources

Sandbox states

View a deployment

Deployment overview

Sandbox configuration

Access a deployed model

Find the endpoint address

Test the deployment

Clone a deployment

Stop a deployment

Billing and usage