Skip to main content
Model Distillation exposes two public HTTP surfaces. Both authenticate with a W&B API key, but they have different responsibilities and tenancy controls.

​
Management API

Studio uses this same API. Resources are durable and asynchronous operations usually return 202 Accepted with a queued, building, running, or pending state. Poll the resource’s GET endpoint rather than holding the request open. The optional Wandb-Entity header selects an accessible team. Resource paths and bodies do not repeat entity ownership.

​
Inference proxy

Use an OpenAI SDK with the proxy base URL. The model field selects a task alias/version or {provider}/{model}. Team tenancy is selected through metadata["wandb.entity"] in the request body so it works consistently across OpenAI SDKs.

​
Interactive endpoint pages

The endpoint pages in this section are generated from the checked-in OpenAPI 3.1 specifications. They include request builders, schemas, response codes, and copyable examples for both surfaces.

Management API

Manage providers, tasks, datasets, fine-tunes, evaluations, routing, and Automation.

Inference proxy

Send OpenAI-compatible Chat Completions requests through tasks or registered providers.
Agents can start with llms.txt and read the raw Management OpenAPI or Inference proxy OpenAPI specifications directly.
The playground sends real requests. Creating datasets, fine-tunes, evaluations, and Automation can consume inference or training resources; routing writes can change production traffic.
Last modified on August 25, 2026