Qwen 3.5 Jinja Template – Restores Qwen /no_thinking behavior!

Posted by Substantial_Swan_144@reddit | LocalLLaMA | View on Reddit | 14 comments

Hi, everyone, As you know, there is no easy way to restore Qwen's thinking behavior in LMStudio. Qwen allows --chat-template-kwargs '{"enable\_thinking": false}', but there is no place there to turn this behavior on and off, like with old models. Therefore, I have created a Jinja script which restores the behavior of the system flag prompt /no\_thinking. That is, if you type /no\_thinking in the system prompt, thinking will be disabled. If omitted, it will be turned on again. The downside: in more complicated problems, the model may still resort to some thinking when responding, but it's not as intense as the overthinking caused by the regular thinking process. Please find the template here: [https://pastebin.com/4wZPFui9](https://pastebin.com/4wZPFui9)

Reply to Post

14 Comments

[-]

Addyad@reddit

[https://github.com/Addy-ad/AIstuff/tree/main/lms](https://github.com/Addy-ad/AIstuff/tree/main/lms) I created a batch script to read your local models and add "enable\_thinking" and "truncate\_thinking" toggles. It works by creating a custom model.yaml file as an alias for whichever model you select. When you run the script, it lists your local models, you pick a number from the list, and then you give it a virtual or alternative name. The script then generates that model.yaml in the custYAML folder. Just make sure to edit your jinja template so the model actually knows how to use these new toggles.

[-]

manfred_exz@reddit

Brilliant! I've been looking for a way to toggle thinking for those model. Works very well.

[-]

Gohab2001@reddit

Qwen 3.5 supports dynamic reasoning, meaning you can easily switch its thinking mode on or off. LM Studio actually has a built-in UI toggle for this, but it only shows up for official `lmstudio-community` models since they come pre-packaged with a [model.yaml](https://lmstudio.ai/docs/app/modelyaml) file that activates the button. You don't need to redownload your GGUF to get the toggle, though. You can just jerry-rig your current setup. Head over to `\.lmstudio\hub\models`, create a folder for the repo owner (like `unsloth`), and inside that, make a subfolder for the model (like `qwen3.5-27b`). Just create a `model.yaml` file in there and paste this block into it: [https://pastebin.com/HDt34yA8](https://pastebin.com/HDt34yA8)

[-]

jcmyang@reddit

Thanks for the tip. After creating the model.yaml I can toggle reasoning on and off. However, I am running into a problem where I get two model entries in the list of LLMs - one for the original name (qwen3.5-27b) and another one with the new name (unsloth/qwen3.5-27b). Is there a way to get rid of the old name?

[-]

Zealousideal_Lie_850@reddit

Just checked your code, you don't need the "ns\_flags" - just set the "enable\_thinking" variable directly: {# ================= auto-disable thinking via system flag ================= #} {%- if messages and messages[0].role == 'system' %} {%- set sys_text_check = render_content(messages[0].content, false, true) | string %} {%- if '/no_thinking' in sys_text_check %} {%- set enable_thinking = false %} {%- endif %} {%- endif %} {# ========================================================================== #}

[-]

FluoroquinolonesKill@reddit

In llama.cpp, can’t you just do --reasoning-budget 0? That’s what I did. Seems to work fine.

[-]

Substantial_Swan_144@reddit (OP)

This is for LMStudio, not LLama.cpp. Unfortunately, there's no --reasoning-budget flag there.

[-]

darkavenger772@reddit

You can add to the jinja template on the model settings page before you load it (very bottom) and it disables the thinking for the model altogether {%- set enable\_thinking = false %}

[-]

Substantial_Swan_144@reddit (OP)

But it doesn't make the flag toggable. This version is.

[-]

Pristine-Woodpecker@reddit

>but there is no place there to turn this behavior on and off, like with old models. What? LM Studio is definitely able to toggle thinking on and off for models with that template parameter. Maybe it just needs an update for Qwen 3.5.

[-]

Substantial_Swan_144@reddit (OP)

Qwen 3.5 ignores this parameter. Maybe the community models have this setting enabled, but the template allows you to toggle thinking for all Qwen 3.5 models, not just community models.

[-]

Pristine-Woodpecker@reddit

I don't understand what you are saying at all. Thinking enabled or disabled, in all modern models (i.e. not the original Qwen3), is controlled by template parameters - that typically cause empty <think></think> blocks to be injected. LM Studio already supports flipping that template parameter with a UI toggle. I've used this with Magistral for example. What I don't know is if they have generic support for this or hardcoded it for a few models.

[-]

Substantial_Swan_144@reddit (OP)

The graphical interface has no way to toggle thinking specifically for Qwen3.5, nor is /no\_think or /no\_thinking supported by defualt. This template restores it.

[-]

Skyline34rGt@reddit

From LmStudio Discord solution is make yaml file - [https://lmstudio.ai/docs/app/modelyaml](https://lmstudio.ai/docs/app/modelyaml) and put it to C:\\Users\\xyz.lmstudio\\hub\\models\\qwen\\qwen35b How? Idk. I just use Community LmStudio model which have thinking toggle.