add qwen 3 chat template post

Author: cfahlgren1

_blog.yml

@@ -5921,4 +5921,12 @@
     - hub
     - inference
 
-
+- local: qwen-3-chat-template-deep-dive
+  title: "The 4 Things Qwen-3's Chat Template Teaches Us"
+  author: cfahlgren1
+  thumbnail: /blog/assets/qwen-3-chat-template-deep-dive/thumbnail.png
+  date: April 29, 2025
+  tags:
+    - qwen
+    - chat-template
+    - llms

assets/qwen-3-chat-template-deep-dive/thumbnail.png

@@ -0,0 +1,89 @@
+---
+title: "The 4 Things Qwen-3’s Chat Template Teaches Us"
+thumbnail: /blog/assets/qwen-3-chat-template-deep-dive/thumbnail.jpg
+authors:
+- user: cfahlgren1
+---
+
+# The 4 Things Qwen-3’s Chat Template Teaches Us
+
+_**What a boring Jinja snippet tells us about the new Qwen-3 model.**_
+
+The new Qwen-3 model by [Qwen](https://huggingface.co/qwen) ships with a much more sophisticated chat template than it's predecessors Qwen-2.5 and QwQ. By taking a look at the differences in the Jinja template, we can find interesting insights into the new model.
+
+<h2 style="text-align: center; margin-bottom: 0.5rem; font-style: italic;">Chat Templates</h2>
+<ul style="display: flex; justify-content: center; list-style: none; padding: 0; margin: 0;">
+  <li style="margin-right: 1rem;"><a href="https://huggingface.co/Qwen/Qwen3-235B-A22B?chat_template=default">Qwen-3 Chat Template</a></li>
+  <li style="margin-right: 1rem;"><a href="https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct?chat_template=default">Qwen-2.5 Chat Template</a></li>
+  <li><a href="https://huggingface.co/Qwen/QwQ-32B?chat_template=default">Qwen-QwQ Chat Template</a></li>
+</ul>
+
+## 1. Reasoning doesn't have to be forced
+
+_**and you can do it via a simple prefill...**_
+
+Qwen-3 is unique in it's ability to toggle reasoning via the `enable_thinking` flag. When set to false, the template inserts an empty <think></think> pair, telling the model to skip step‑by‑step thoughts. Earlier models baked the <think> tag into every generation, forcing chain‑of‑thought whether you wanted it or not.
+
+```jinja
+{# Qwen-3 #}
+{%- if enable_thinking is defined and enable_thinking is false %}
+    {{- '<think>\n\n</think>\n\n' }}
+{%- endif %}
+```
+
+QwQ for example, forces reasoning in every conversation.
+
+```jinja
+{# QwQ #}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n<think>\n' }}
+{%- endif %}
+```
+
+If the `enable_thinking` is true, the model is able to decide whether to think or not. 
+_More often than not it will think._
+
+## 2. Context Management Should be Dynamic
+
+_Qwen-3 utilizes a rolling checkpoint system, intelligently preserving or pruning reasoning blocks to maintain relevant context. Older models discarded reasoning prematurely to save tokens._ 
+
+Qwen-3 introduces a "**_rolling checkpoint_**" by traversing the message list in reverse to find the latest user turn that wasn’t a tool echo. For any assistant replies after that index it keeps the full `<think>` blocks; everything earlier is stripped out.
+
+**Why this matters**:
+- Keeps the active plan visible during a multi‑step tool call.
+- Supports nested tool workflows without losing context.
+- Saves tokens by pruning thoughts the model no longer needs.
+- Prevents "stale" reasoning from bleeding into new tasks.
+
+### Example
+
+Here's an example of chain-of-thought preservation through tool calls with Qwen-3 and QwQ.
+![image/png](https://cdn-uploads.huggingface.co/production/uploads/648a374f00f7a3374ee64b99/7OWKkRuO9Qc2L48LYjxVf.png)
+<p style="text-align:center; font-style:italic; font-size:medium;">
+  Check out <a href="https://www.npmjs.com/package/@huggingface/jinja">@huggingface/jinja</a> for testing out the chat templates
+</p>
+
+## 3. Tool Arguments Need Better Serialization
+
+Before, every `tool_call.arguments` field was piped through ` | tojson`, even if it was already a JSON‑encoded string—risking double‑escaping. Qwen‑3 checks the type first and only serializes when necessary.
+
+```jinja
+{# Qwen3 #}
+{%- if tool_call.arguments is string %}
+    {{- tool_call.arguments }}
+{%- else %}
+    {{- tool_call.arguments | tojson }}
+{%- endif %}
+```
+
+## 4. Default Prompts Should be Optional
+
+Qwen‑2.5 automatically inserted a default Alibaba system prompt:
+
+> You are Qwen, created by Alibaba Cloud. You are a helpful assistant.
+
+Unlike Qwen-2.5, which automatically inserted a default Alibaba system prompt, Qwen-3 (alongside QwQ) omits any default prompts, allowing developers full control over the model's persona.
+
+## Conclusion
+
+Qwen-3 shows us that through the `chat_template` we can provide better flexibility, smarter context handling, and improved tool interaction. These improvements not only improve capabilities, but also make agentic workflows more reliable and efficent.