An API for everyoneUser guide
Website ↗
User guideAfter connecting

Adjust thinking effort

First declare the model’s available thinking levels, then choose a level in chat. The two steps are shown below.

Enable the available levels

06Enable thinking levels for the model

  1. Open the configuration file

    This version has no graphical switch for effort levels. Click 打开配置文件 at the top right of settings and open settings.yaml in a text editor.

  2. Find the model you just added

    Under your My API provider, find deepseek-v4.1-flash-1017 and add reasoningEfforts as shown below. Do not edit a model with the same ID under another provider.

  3. Save and check the menu

    Save, then reopen the model menu. Restart Harness if it has not updated. If your version has Model options → Reasoning levels, you can configure the levels there.

Enable thinking levels for the model
Chinese interface · Open full-size screenshot

Add reasoningEfforts only within the model entry you just created. Keep the existing address, credential reference, and other settings. This is a partial example: do not replace the whole file with it.

models:
  - id: deepseek-v4.1-flash-1017
    reasoningEfforts:
      low: low
      high: high
      max: max

Leading spaces matter. Indent with spaces and keep low, high, and max spelled exactly as shown.

Choose a thinking level

09Choose the thinking level

  1. Open the model menu again

    With the My API model selected, click its name at the bottom right and open Reasoning levels (推理等级).

  2. Choose Low, High, or Max

    Start with High for everyday tasks; try Max for harder tasks and compare the results and usage on your own work.

  3. Check the selected level

    After choosing High, High appears beside the model name. DeepSeek Flash in the image is a display name for the same model.

Choose the thinking level
Chinese interface · Open full-size screenshot
After you select High, the level appears beside the model name.

After you select High, the level appears beside the model name.

Displaying the menu and upstream handling of each level are separate checks. Tested Off, Low, High, and Max requests were accepted; the Off sample returned no reasoning usage. Higher levels may take longer and use more tokens, without guaranteeing a better answer.