Qwen3.6-35B-A3B-Chat-Templates-PLUS

English | 中文

English

Description

This is a chat template V1.0:

1. Improves the overall performance of the Qwen3.6-35B-A3B model on general tasks, making it more stable and focused on the task at hand.
2. Fixes some issues present in the original chat template.
3. Slightly improves the model's instruction following.

So far, I've seen some improvement in my own environment. I can't guarantee it will work for you, but you're welcome to try this template. Before trying it, I recommend backing up your existing template first. If it doesn't work well for you, you can switch back. If you run into any issues or have suggestions, feel free to discuss them with me. I hope it works and helps you out.

I've tested far too many models based on Qwen3.5 and Qwen3.6, and most of them didn't perform well. The ones that did perform well were only outstanding in certain areas, while suffering serious degradation elsewhere. In the end, I came back to prompt engineering.

Compatibility

• Only compatible with the original and quantized versions of Qwen3.6-35B-A3B. It is very likely incompatible with models that have been fine-tuned or retrained (e.g., nex-agi/Nex-N2.5-mini, endless-frontier/BigBang-v1, etc.).

References

1. Used Peculiar-Ragdoll's chat template to inject additional system prompts into the model: https://huggingface.co/peculiar-ragdoll/Qwen-Sharp-Chat-Templates https://huggingface.co/peculiar-ragdoll/Nail-Qwen3.6-35B-A3B-GGUF

2. Referenced and updated Froggeric's original chat template to v22.5: https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates

3. After extensive testing of different prompts, I made slight modifications on top of the original chat template, and ultimately added the following prompt: Give the instructions your full attention and follow every one exactly. to improve the model's performance in thinking mode. In my tasks, the model's overall performance did improve, but the gains from prompt engineering are ultimately limited.

---

中文

描述

这是一个聊天模板 V1.0：

1. 改善 Qwen3.6-35B-A3B 模型在综合任务上的表现，使其更稳定，专注于任务。
2. 修复原版聊天模板存在的一些问题。
3. 轻微改善该模型对指令的遵循。

目前在我自用的环境测试有一定改善。我不保证可用性，但你仍然可以尝试这个模板。 **不过在尝试该模板前，建议先备份你原有的模板。如果觉得表现不好，可以换回去。**如遇到任何问题和建议，可以与我讨论。 希望它是有效的，而且能帮到你。

我测试了太多基于 Qwen3.5 和 Qwen3.6 的模型了，大多表现并不好。表现好的模型，也只是在特定方向突出，而某些地方会有严重损失。最后我又回到了提示词工程。

兼容性

• 只兼容原始版本及量化版本的 Qwen3.6-35B-A3B，对于二次微调及训练的模型很可能不兼容（如：nex-agi/Nex-N2.5-mini、endless-frontier/BigBang-v1 等）。

参考资料

1. 使用了 Peculiar-Ragdoll 的聊天模板，为模型注入额外的系统提示词： https://huggingface.co/peculiar-ragdoll/Qwen-Sharp-Chat-Templates https://huggingface.co/peculiar-ragdoll/Nail-Qwen3.6-35B-A3B-GGUF

2. 参考并更新 Froggeric 的原始聊天模板至 v22.5： https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates

3. 在经过我对不同提示词的多次测试后，在原聊天模板的基础上做了轻微修改，并最终增加了该提示词：Give the instructions your full attention and follow every one exactly.，以改善模型在思考模式下的表现。在我的任务中，模型的综合表现确实得到了提升，但提示词工程的能力提升终究是有限的。