Model routing - CoreWeave Docs

The first slash in model distinguishes a direct provider reference from a proxy model name.

Proxy model names

{"model": "ticket-classifier"}

The proxy reads routing for version 1, chooses among targets with positive weights, replaces model, and calls the target’s provider.

Version pins

{"model": "ticket-classifier@v3"}

A pin resolves exactly version 3. It never falls forward to the latest version and never changes version 1. This makes a new version safe to test before application rollout.

Direct provider calls

{"model": "openai/gpt-5.6-sol?reasoning_effort=medium"}

The prefix before the first slash selects a provider in the resolved W&B entity. The remainder is the configured provider model ID; additional slashes are allowed. Direct calls are useful for relabeling, judges, ad hoc comparisons, and validating provider setup. They bypass project routing, aren’t attributed to any project, and aren’t traced unless the request sets wandb.project. See Metadata and traces.

Entity resolution

The bearer key’s default entity is used unless the body contains:

{
  "metadata": {
    "wandb.entity": "your-team"
  }
}

The requested value must be the caller’s personal/default entity or one of their W&B teams.

Sticky weighted selection

Provide a conversation ID to keep all calls in one conversation on the same target for a routing revision:

{
  "metadata": {
    "wandb.thread_id": "conversation-4827"
  }
}

Changing the routing revision can intentionally move that key to a different target.

API: Call a pinned project version (POST /chat/completions)

Use alias@vN in model. A direct provider call uses provider/model-id instead.

curl --request POST \
  --url "https://proxy.training.wandb.ai/v1/chat/completions" \
  --header "Authorization: Bearer $WANDB_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "model": "ticket-classifier@v2",
    "messages": [{"role": "user", "content": "Classify this request"}],
    "metadata": {"wandb.entity": "your-team"}
  }'

See Create a routed chat completion for the complete schema.

Last modified on September 4, 2026

​Proxy model names

​Version pins

​Direct provider calls

​Entity resolution

​Sticky weighted selection

Proxy model names

Version pins

Direct provider calls

Entity resolution

Sticky weighted selection