Non-reasoning mode
Hi
Thank you for your work, much appreciated.
Do I understand correctly that if I turn off the reasoning mode, there won't be much improvement over the base model?
The issue I'm having is that in agentic workflows, the tools already force the model to reason by issuing prompts like "create step by step plan" "research methods to perform X", and, in my experience, turning on reasoning doubles the time to achieve the goal without noticeably improving the quality.
What's your opinion on this?
Hi,
I think your interpretation is mostly correct.
In agentic workflows, the tools already push the model into structured reasoning through their prompts. So in those cases, turning on reasoning mode really comes down to whether the extra overhead is worth it. But when a task doesn’t have a clear structure, needs the model to plan on its own, or involves longer reasoning chains, reasoning mode can still help with stability and consistency.
It’s just that these benefits are much less noticeable when the whole process is tool‑driven.
Thanks for the explanation.
So would you say your finetuning would have a chance of measurably improving the output quality in non-reasoning mode, or it doesn't work that way?
sorry if it's a stupid question, I'm but a simple user, looking to automate some mundane daily tasks with agents on my own hardware. Qwen 3.5 was a real game changer for my case, what we had before was absolutely not suitable for anything beyond "hello world".
How do you turn off reasoning?
I tried /no_think, which doesn't work anymore with qwen3.5.
the "chat_template_kwargs": {"enable_thinking": false} also didn't work for me.
adding {%- set enable_thinking = false %} to the template also didn't work.
(full disclosure: I have no idea what I'm doing, just trying thing I saw people doing on the internets.)
How do you turn off reasoning?
I tried
/no_think, which doesn't work anymore with qwen3.5.the
"chat_template_kwargs": {"enable_thinking": false}also didn't work for me.adding
{%- set enable_thinking = false %}to the template also didn't work.(full disclosure: I have no idea what I'm doing, just trying thing I saw people doing on the internets.)
if you are using llama.cpp to run this model, you can try "--reasoning-budget 0" to trun off thinking mode.
How do you turn off reasoning?
I tried
/no_think, which doesn't work anymore with qwen3.5.the
"chat_template_kwargs": {"enable_thinking": false}also didn't work for me.adding
{%- set enable_thinking = false %}to the template also didn't work.(full disclosure: I have no idea what I'm doing, just trying thing I saw people doing on the internets.)
if you are using llama.cpp to run this model, you can try "--reasoning-budget 0" to trun off thinking mode.
thanks, --reasoning-budget 0 this param it work for me
How do you turn off reasoning?
I tried
/no_think, which doesn't work anymore with qwen3.5.the
"chat_template_kwargs": {"enable_thinking": false}also didn't work for me.adding
{%- set enable_thinking = false %}to the template also didn't work.(full disclosure: I have no idea what I'm doing, just trying thing I saw people doing on the internets.)
if you are using llama.cpp to run this model, you can try "--reasoning-budget 0" to trun off thinking mode.
"--reasoning off" reasoning budget may have some issues for some models, but if it works keep it that way.