A support model may need restrained replies while a brainstorming model benefits from more variation. Open WebUI model parameters store generation defaults on a workspace model, so each model-picker entry can use its own sampling settings.

A workspace model wraps a connected base model and stores its own parameter overrides. Instance defaults apply first, then the model's defaults; account or chat settings can override ordinary generation parameters for a request. The inference provider still determines which parameters and ranges it supports.

Editing requires ownership of the workspace model or administrator access. A change to a shared model affects other people using it, so a private preset is a better starting point for testing; its base model must already answer chat requests.

Steps to tune Open WebUI model parameters:

  1. Open Workspace from the sidebar.

    The Models tab opens by default; if another workspace tab is active, select Models.

  2. Click the workspace model's name to open its editor.

    Support Triage is a private preset used here.
    Related: How to create a model preset in Open WebUI

  3. Confirm that Base Model (From) names the intended connected model.

    The base model is company-chat here. A missing or unavailable base model requires a working provider connection before tuning.
    Related: How to refresh the Open WebUI model list

  4. Expand Advanced Params.
  5. Record the current Temperature setting for rollback.

    Default means the workspace model has no explicit override for this field. Restoring Default later removes the model-level override; other applicable defaults can still supply a value.

  6. Click Default beside Temperature to switch it to Custom.
  7. Set the Temperature number field to 0.2.

    A lower temperature generally reduces sampling variation. Changing only this parameter makes comparisons easier; top_p is another sampling control, while max_tokens limits output tokens rather than words. Unsupported parameters may be ignored or rejected by the provider.

  8. Click Save & Update at the bottom of the editor.
  9. Reopen the same workspace model's editor.
  10. Expand Advanced Params again.
  11. Confirm that Temperature remains Custom at 0.2.
  12. Start a New Chat.
  13. Open the model picker in the chat composer.
  14. Select the tuned workspace model.

    A fresh chat without account, chat, or model-control parameter overrides isolates the saved model default.

  15. Enter a short test question in the chat composer.
    A customer cannot sign in. What should I check first?

    A synthetic question avoids sending real customer information during testing.

  16. Click the upward-arrow Send message button.
  17. Confirm that the tuned workspace model returns a reply without an error.

    The reopened editor proves persistence, and the reply shows that the model remains usable. Exact forwarding requires the provider's request log; a model's description of its own parameters cannot prove the values it received. Reply wording alone also cannot establish that the provider honored a sampling setting.

Share this guide