> ## Documentation Index
> Fetch the complete documentation index at: https://docs.coreweave.com/llms.txt
> Use this file to discover all available pages before exploring further.

# How Model Distillation works

> See how real product usage becomes a trained, evaluated, and safely deployed model.

Model Distillation adds an improvement loop around an AI feature you already have. Your application keeps serving requests while training and evaluation happen separately in the background.

```mermaid theme={"system"}
flowchart LR
  App[Your application] --> Use[Real product usage]
  Use --> Data[Training dataset]
  Data --> Train[Candidate models]
  Train --> Compare[Quality comparison]
  Compare --> Deploy[Gradual deployment]
  Deploy --> App
```

## 1. Connect an existing feature

Your application calls Model Distillation through the familiar Chat Completions interface. Give the feature a stable task name, and that name remains unchanged as the model behind it improves.

## 2. Learn from real examples

Model Distillation collects the inputs and successful answers your application already produces. You choose which examples represent the behavior you want the new model to learn.

If the existing answers are not ideal, you can use a stronger model to create improved training answers before training.

## 3. Train and compare candidates

Train one model or a sweep of different models. Every candidate is tested on the same examples so you can see whether it preserves the quality of the model you use today.

## 4. Deploy without a risky switch

Send a small amount of traffic to the winner first, or deploy it fully after evaluation. The previous model can remain available for a quick rollback.

## 5. Repeat automatically

Automation can wait for enough new examples, prepare the next dataset, train several candidates, compare them, and deploy the best model that clears your threshold.

## What you control

* which product behavior becomes a task;
* which examples and answers the model learns from;
* which models are trained and compared;
* how good a candidate must be before deployment;
* how much traffic the winner receives.

<Card title="Follow the complete workflow" href="/model-distillation/quickstart" arrow="true">
  Connect a task and take it from real usage to a deployed model.
</Card>

<Accordion title="API: Workflow endpoints">
  An agent can perform the same loop through these public resources:

  1. `PUT /tasks/{alias}` and `POST /chat/completions` to connect traffic;
  2. `POST /tasks/{alias}/datasets` to freeze training and validation data;
  3. `POST /tasks/{alias}/finetunes` for each candidate;
  4. `POST /evals` to compare the candidates;
  5. `PUT /tasks/{alias}/versions/{version}/routing` to deploy a winner.

  Use the [Management OpenAPI specification](/openapi/model-distillation/management.openapi.yaml) and [Inference proxy OpenAPI specification](/openapi/model-distillation/proxy.openapi.yaml) for exact schemas.
</Accordion>