Skip to main content
Before you can benchmark a model or call it from an agent, it needs to be running: loaded onto a GPU and exposed as a live inference endpoint. This page walks through deploying a model from the Eval Models area, watching it come online, and confirming it is ready to serve requests.
Run Inference here vs. in TrainThis page covers deploying a model as a persistent, OpenAI-compatible endpoint inside Eval. If you only want a quick chat sanity-check against a freshly trained adapter, use Train → Run Inference instead.

​
Prerequisites

  • A model available in your workspace: uploaded weights, a model from Train, or a published model. See Add a model.
  • At least one GPU node with free capacity in your environment.

​
Steps

​
1. Open the Models page

In the Eval tab, click Models in the left sidebar. The AI Models page lists every model in your workspace, both base models and deployed endpoints. Use the filter chips (All, Running, Deployed, Base, Public) or the search box to find the model you want to deploy. The AI Models list in Eval, showing deployed models and filter chips

​
2. Open the model card

Click a model to open its Model Card. The card shows the model’s details, benchmark leaderboards, and, in the panel on the right, its endpoint and basic information. To deploy it for inference, click Run Inference in the top right corner. The gemma4 model card with the Run Inference button in the top right

​
3. Configure and start inference

The Run Inference dialog opens. At the top, the Will be served as line shows the identifier your endpoint will use, for example demoaccount/gemma4. Pick a GPU from the dropdown, which lists nodes that still have free capacity such as 1/1 free, then review the inference configuration: When you’re happy with the settings, click Run model. The Run Inference dialog with GPU selection and inference configuration fields
The defaults are a good starting point. For benchmarks that expect long answers, like detailed reasoning or code, raise Max tokens and Context length. For deterministic scoring, lower Temperature.

​
4. Monitor deployment status

After you click Run model, the model begins deploying. A banner reads “Deploying… the model is starting up”, the status badge switches to Deploying, and a Stop Model button appears in the top right. The panel on the right updates with the new endpoint details: its LiteLLM Name, Endpoint URL, access Token, and a Status of deploying. New Logs and Usage tabs also appear. The Usage tab is where you track requests, tokens, and latency once the model is serving traffic. See Monitor model usage. The deploying banner and status badge while the endpoint starts up

​
5. Inspect the deployment logs

Open the Logs tab to watch the deployment in real time. The logs stream the runtime setup as the Ray cluster connects, the serve application starts, and the model weights load. Deployment logs streaming during startup Loading can take a minute or two depending on model size. Click Refresh if you want to pull the latest lines manually.

​
6. Verify the model is running

When startup finishes, the logs report that the application is ready. For example:
The Status badge turns to running (green). Your model is now live and serving requests. Logs showing the application is ready and the status set to running Back on the Models page, the model now appears under a Running section, and the Running filter count goes up by one. The Models list with the newly running model in the Running section

​
Accessing the inference endpoint

Once the model is running, the panel on the right of the model card gives you everything you need to call it: The endpoint is OpenAI-compatible, so you can call it with any OpenAI client or a plain curl request:
A running model holds a GPU for as long as it stays deployed. When you’re done, open the model card and click Stop Model to free the resources.

​
Next Steps

Last modified on July 14, 2026