Skip to main content
Applies to:
  • Plan -
  • Deployment -

​
Summary

Manual Vertex model entries return 404 or permission errors after you disable the default registry. Likely causes: per-model location or model format mismatch, or region/global behavior for multi-region models (e.g., gemini-3.6-flash). Fix by matching the default registry settings: exact model ID, correct format, and the model location (or use a region-specific API base URL).

​
What is happening

When Braintrust includes the default Vertex registry some models are registered with a per-model location of global. Manually adding the same model without the global location (or with the wrong format) can cause Vertex to look in a different region and return NOT_FOUND (404). Multi-region/global models (like Gemini 3.6 Flash) require either the global model location or a region-specific API base URL to keep traffic in a specific geographic boundary. Permission errors indicate the Vertex-calling service account lacks the necessary role on the target project.

​
Fix or suggestion

​
Option 1: Match default registry per-model settings

  1. Use the exact Vertex model name. Example:
    • publishers/google/models/gemini-3.6-flash
  2. Set the model format to the provider-appropriate value (Gemini when set in the UI / google when set with the API), not the default openai.
  3. Set the model location to global for models that the default registry marks as global.
  4. Save and retry the call.

​
Option 2: Force US-only traffic for global/multi-region models

  1. Keep the provider location as your preferred region (e.g., us-west1) if needed.
  2. Change the provider API base URL to the US Vertex endpoint to ensure requests stay within the US for global models (or use the EU API URL if you need that):
    • Example: https://aiplatform.us.rep.googleapis.com
  3. Test this change in a custom provider first, then apply to the live provider if successful.

​
Check service account permissions (if errors persist)

  • Ensure the Vertex-calling service account has roles that allow model invocation (for example: AI Platform Developer or equivalent) on the project hosting the model.

​
How to confirm it worked

  • Run a quick Playground or test prompt that calls the model. Successful calls return a model response instead of NOT_FOUND or permission errors.
  • Verify provider/model configuration shows the expected model name, format, and location (or that the provider API base URL is set to the US endpoint).

​
Notes

  • 404/NOT_FOUND usually means wrong model name or mismatched region/location. Permission errors mean the service account needs the correct IAM role.
  • Google model location docs