Serving

BETA

This is a beta feature. For more information and access requests, please contact us through the Support Portal.

Serving puts a finished fine-tuning adapter online so you can query it like any other model in the project. After training succeeds, the adapter is listed in Inference → Models, tagged as fine-tuned. It is not available for inference until you serve it.

Shared base models in the catalog are already available. This step applies only to adapters you trained.

You can tick Serve automatically after training when you create the job, or serve later from job details or the model card. Calling the adapter uses a project API key, the same as the rest of Studio; see Projects.

What this enables

Use serving when a fine-tuning job has succeeded and you want the adapter online for inference.

From the job or the model catalog you can:

• Serve an adapter so Playground and API requests succeed
• Leave it registered but unloaded until you need it
• Unload it when you no longer need it online, without deleting it
• Copy Python and cURL examples from the model detail view once it is serving

How it works

1. Training finishes. The adapter is registered in Inference → Models. Status is Not serving unless you enabled auto-serve on the job.
2. Serve the adapter. Click Serve model on the job, or open the fine-tuned model in the catalog and serve it there.
3. Wait through Queued. Status becomes Queued until serve finishes. This may take a moment. You can leave the page; the catalog updates when it completes.
4. Use it. When status is Serving, pick the model in the Playground or call the model id from the card with a project API key. Capabilities (context length, tools, reasoning) follow the base model.
5. Unload when you are done. Unload takes the adapter offline. Status returns to Queued, then Not serving. The adapter stays in the catalog and can be served again later.

---

Adapter status

Getting to a serving or unloaded state takes up to one minute. The catalog shows one of three statuses while that work runs. The job details page also shows Adapter (Not serving or Serving) next to job status.

Status	Meaning	What you can do
Not serving	The adapter is registered. It is not loaded, so inference requests to it fail.	Serve model, or Delete if you no longer need the adapter.
Queued	A serve or unload request was accepted and is still running.	Wait until status becomes Serving or Not serving. Do not start another serve or unload on the same adapter.
Serving	The adapter is loaded. Project API keys can call it.	Unload, use the Playground, or call the API.

Not serving

The adapter is in the catalog but not loaded. Open the model and click Serve model, or serve from the finished job.

Not serving — the adapter is listed but not loaded for inference.

Queued

After you click Serve (or Unload), status becomes Queued until the change finishes. This may take a moment. If it stays Queued for a long time, contact us through the Support Portal.

Queued — serve or unload is in progress.

Fine-tuning job progress (download dataset, train, upload adapter) is separate and stays on the job page. Catalog status is only whether that finished adapter is loaded for inference.

Serving

The adapter is loaded. Use the model id shown on the card in Chat Completions or the Playground. The model detail view includes Python (OpenAI) and cURL examples. Requests appear in Inference → Activity log, scoped to the project.

Serving — copy the API example to call it, or Unload to take it offline.

---

In the product

Feature	What you get
Job details	After a successful run, Serve model and an Adapter field (Not serving or Serving).
Model catalog	Fine-tuned adapters listed with shared models. Open a card to serve, unload, or delete.
Setup examples	Once Serving, Python (OpenAI) and cURL snippets in the model detail view.
Activity log	Inference traffic to the adapter after it is serving, scoped to the project.

When the adapter is ready, continue in the Playground or Chat Completions.