Make reasoning effort steering more intuitive in the Jinja template
I tested your Jinja template (based on froggeric) against another one from Reddit. Yours works much better because it formats the thoughts correctly and prevents infinite overthinking from eating up all available tokens.
However, I found the reasoning effort steering confusing for LM Studio users. Changing the thinking mode inside the template is not trivial, at least for me. I tried changing the line
{%- set _initial_effort = 'medium' %}
with
{%- set _initial_effort = 'xhigh' %}
and it did not give the intended result.
Paradoxically, setting a reasoning budget in LM Studio seems to force the effort to low, ignoring the value completely. I find it also awkward to manually type tags like <|think_xhigh|> in the chat prompt.
Could you make steering more trivial in the documentation, or directly within the Jinja template configuration ?
Nevertheless, I realized that your default medium mode is actually the best fit for me, giving a deep analysis without losing the output.
One quick note on settings: for the template to shine with Qwen 3.8, I noticed it helps to stick to the official recommendations (Temp 1.0, Top-K 20, Top-P 0.95) and disable repeat penalty, presence penalty, and min-p.
Thanks for your work.
Thank's a lot for the feedback and the heads up! That's not very user friendly indeed. Templates take reasoning effort through the chat template args or inline tags in prompt, not through OpenAI API effort parameter, but you can set it through passing template args without changing the template itself :) I do not recommend LM Studio either way: oMLX and llama.cpp are my go-to backends. For this template, the default thinking mode it sets is the tested and optimal setting. The system prompt itself counteracts overthinking by inducing convergent thinking, even if it does not target it. So even though the models get much more efficient at medium thinking effort in structured problems, it can be hard to get it to think freely, deeply, and divergently like it does under default template and xhigh effort. I'll think about how to address all these things.
another one from Reddit.
Please link it.
If you're interested, here's the other Jinja template on Reddit. But it seems that Froggeric has already incorporated what he liked the best from it.
https://www.reddit.com/r/LocalLLaMA/s/4Ldf9eok7F