Setting Qwen3.8 reasoning effort with Ollama

[pjv]

I've been through all the OMP docs and all the Ollama docs and all the search engine results that i can find but I cannot configure OMP so that when I cycle through the thinking levels via shift-tab in the TUI that it seems to actually have any effect on the qwen3.8:27b-mtp-q8_0 model I have running on an ollama server on my LAN (not the same host where OMP is running).

I have the provider configured like this:

  ollama:
    baseUrl: http://ai.local:11434/v1
    api: openai-completions
    auth: none

My currently un-successful attempt at configuring the model looks like this:

models:
      - id: qwen3.8:27b-mtp-q8_0
        reasoning: true
        contextWindow: 150000
        maxTokens: 8192
        input:
          - text
          - image
        compat:
          supportsReasoningEffort: true
          thinkingFormat: openai
          reasoningEffortMap:
            off: none
            low: low
            medium: medium
            xhigh: xhigh

A related side question: in the TUI, when you hit shift-tab it cycles through these thinking levels: off, auto, low, med, high, max and I have no idea how those are specifically supposed to map to either reasoningEffortMap or thinkingLevelMap in the models.yml config. In Pi, when you configure a thinkingLevelMap whatever levels you put in there that aren't mapped to null is what you get when you cycle the level in the TUI, but this does not seem to be the same behavior in OMP. No matter what I put in my config, the TUI seems to cycle through the same levels (off, auto, low, med, high, max).

Am I missing some detailed thinking level docs somewhere?

[pjv]

Here's what I've discovered through watching an agent massively overthink through prompts to diagnose the above problem. This config seems to be working:

models:
      - id: qwen3.8:27b-mtp-q8_0
        reasoning: true
        thinking:
          mode: effort
          efforts:
            - low
            - medium
            - xhigh
        compat:
          supportsReasoningEffort: true
          reasoningDisableMode: none-effort
          thinkingFormat: openai

Most of this thinking related config seems entirely undocumented as far as I could find.

[nbonfire]

I still can't disable reasoning with this config, if I set to "off" i still see the reasoning happen.

[pjv]

That’s interesting. It’s still working correctly for me with Qwen models (lately using qwen3.8-flash-next). Omp v. 18.2.1:

[nbonfire]

thanks, turns out it was PEBKAC - I kept seeing these snippets of models.yml that made me think i could have a root level models key w/o specifying the provider. once I moved it into the ollama provider it worked.