Routing and versions - CoreWeave Docs

Routing defines which model serves a project version. Every target contains a model reference, non-negative weight, and optional parameter overrides.

Weights are relative

Weights don’t need to add to 100. A configuration with weights 3 and 1 sends approximately 75% and 25%. A zero-weight target stays available in the configuration but receives no new weighted traffic. Use zero weight to preserve the previous model for visibility and rollback after a full promotion.

Sticky routing

When a request supplies wandb.thread_id or gen_ai.conversation.id, the proxy deterministically keeps that conversation on one target for the routing revision. Without a conversation ID, each request is selected independently.

Target parameters

To add values such as temperature, reasoning effort, response format, or provider-specific controls, expand Parameters under a model. Numbers, booleans, arrays, and objects are parsed from JSON. Other text remains a string. The stored override wins over the same field supplied by the caller.

Save to an existing or new version

Open the project’s Routing page. Add or edit targets, choose the destination in the version bar, and select Save routing. The version bar is part of the same save action:

Choose an existing version to replace its desired routing.
Choose the new-version option to create a new version. The version is created only when you select Save routing. Call the new version with name@vN.

Each save publishes a new routing revision: a numbered configuration for that version that includes its targets, weights, parameter overrides, and trace-sampling rate. Publishing is asynchronous, so a saved configuration serves traffic only after its deployment status becomes Applied at that revision.

If you delete a version, callers pinned to it receive 404 responses. You can’t delete version 1.

API: Replace routing for a project version (PUT /tasks/{alias}/versions/{version}/routing)

Routing updates replace the whole configuration, so the request must include trace_sampling_rate as well as every target. To keep the existing sampling rate, pass it in the request. If you want the previous model available for rollback, keep it at weight 0:

curl --request PUT \
  --url "https://distillation.training.wandb.ai/v1/tasks/ticket-classifier/versions/2/routing" \
  --header "Authorization: Bearer $WANDB_API_KEY" \
  --header "Wandb-Entity: your-team" \
  --header "Content-Type: application/json" \
  --data '{
    "trace_sampling_rate": 1,
    "targets": [
      {"model_ref": "openai/gpt-5.6-sol", "weight": 0},
      {"model_ref": "wandb-inference/your-fine-tuned-model", "weight": 1}
    ]
  }'

Poll the returned routing state until deployment is applied at the current revision. See Replace routing for a task version.

Last modified on September 21, 2026

​Weights are relative

​Sticky routing

​Target parameters

​Save to an existing or new version

Weights are relative

Sticky routing

Target parameters

Save to an existing or new version