| --- |
| license: apache-2.0 |
| base_model: |
| - Qwen/Qwen3.6-27B |
| - Qwen/Qwen3.6-35B-A3B |
| - Qwen/Qwen3.5-27B |
| - Qwen/Qwen3.5-35B-A3B |
| pipeline_tag: text-generation |
| library_name: transformers |
| tags: |
| - chat-template |
| - jinja |
| - jinja2 |
| - qwen3 |
| - template-fix |
| - bugfix |
| --- |
| ### 23.05 22:00: v1.1.5 is released. Agentic and Tool use works 100% for hours now. Try it out! |
| ### If you use mainly agentic coding you can also test the "additional-systemprompt" variant out. It adds minimalistic Systempromt Enhancements to 1.1.5 that should work in every environment and actively enhancing correct coding creation and behaviour. |
|
|
| TL:DR Download your fixed and finally working Qwen3.5 and Qwen3.6 jinja chat-template here. V1.1.2 with new tool call error bug fixes and additional edge case enhancements is now released. |
|
|
| You can use this template for any Qwen3 model (only Qwen3-Coder has another template format) |
| Compatible with Qwen3, Qwen3.5 and Qwen3.6 in any OpenWeight size yet released! |
|
|
| #### Update: V1.1.5 is out in the wild now! |
| Additional fixes to tool calling have been done, verified correct developer role handling and other improvements. |
|
|
|
|
| #### Preserve Thinking: |
| To enable preserve-thinking (Agent can remember what he thought about in older turns) pass the following parameter to llama.cpp startup: |
| On Linux&MacOS (Windows need slightly different which cant be correctly displayed here...): --chat-template-kwargs '{"preserve_thinking": true}' |
| |
| #### OpenCode: |
| If using OpenCode do notice they have an open bug for showing thinking content as plain text. Open Bug since 4 months(https://github.com/anomalyco/opencode/issues/11439). Meanwhile you can use the llama.cpp Server switch to hide thinking completely(but still let the model think) using the additonal llama.cpp parameter: |
| --reasoning-format deepseek |
| |
| |
| # Welcome to the Template Rebuild Project |
| |
| How do we got here: |
| Annoyed from various bugs with Qwen3.5 and Qwen 3.6 models tool calling etc. I decided to test many different community variants of Qwen Templates, like probably many others out there. |
| At the moment there isnt any completly working chat template for the Qwen3.5 and Qwen3.6 models. |
| |
| Community patches were collected and patched together to made a single final version. But it still contained bugs. |
| |
| In the spirit and effort of froggeric https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/ |
| I want to make this awesome model category useful for any case. |
| |
| So I did intense preresearch about the models nature, collected a huge list of commonly used ai-tools and which requests do they make to LLM-Endpoints. |
| |
| All of this was then fed into a new, from the ground up built version of a chat-template. |
| |
| I believe I did that groundwork, currently I didnt covered all off the over 90 common AI-tools but only like 30, I will continue to enhance the information pile and get 99% tool coverage(100% is probably unnecessary and impossible, considering the fast evolving environment, also many tools use the same calling mechanisms). |
| |
| This will lead to a final version of the file that is useable for any usecase you can think of. |
| |
| Currently a third version is released as v1.1.2 which I would call stable |
| |
| 65 specific tests has been run and actual behaviour also manually verified by me. |
| |
| Please feel free to report any bug you encounter with the template, it will be looked in manually from me and be analyzed and fixed. |
| |
| Have fun! |
