Qwen Sharp Chat Templates

Qwen Sharp Chat Templates

Replaces the chat template of Qwen3.5, 3.6 and 3.8 models so they think less before answering and write tighter replies. For people running Qwen locally for coding and knowledge work.

Made by peculiar-ragdoll on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Qwen Sharp Chat Templates is a replacement chat template for Qwen3.5, 3.6 and 3.8 models, aimed at knowledge work and coding. It builds on froggeric's fixed Qwen templates and adds a system prompt that is always appended. The card says the models then use fewer thinking tokens and trim filler from their replies.

Who it suits: People who run Qwen models on their own machine and want faster, tighter answers without changing the model.

Industry | Who uses it | How they use it | Example
Software | Developers | Swap the stock template in a local Qwen setup | Speed up fixes on repository tasks
AI tooling | Agent builders | Keep tool calls working across runtimes | Avoid duplicate tool instructions in LM Studio
Knowledge work | Analysts | Get shorter replies in long chats | Cut filler in multi-turn conversations

If you only use hosted chat apps, you can skip it.

Where to use it:

• llama.cpp, where template options pass in with a chat template kwargs flag
• LM Studio with the MLX backend, where the template detects the runtime's own tool protocol
• transformers and oMLX, which can pass template options

Use case | Who | What they get back
Coding on repositories | Developers | Fixes reached faster at similar solve counts
Agent runs with tools | Agent builders | Tool calls that parse instead of leaking raw text
Long conversations | Everyday users | Replies with less filler per turn

How to set it up:

• A model from the Qwen 3.5, 3.6 or 3.8 line, running on your machine
• A runtime that lets you replace the stock chat template
• Python with pip, to fetch the files

1. Install the Hugging Face hub tool.
2. Download the template files.
3. Point your runtime at the template in place of the model's stock one.

pip install -U huggingface_hub
hf download peculiar-ragdoll/Qwen-Sharp-Chat-Templates

How to use it: After you swap the template in, chat as normal. The card names an override called suppress_tool_instructions, which turns off the template's own tool block where your runtime can pass template options. The appended prompt can also be switched off per request.

Tool | What it does
suppress_tool_instructions | Overrides the template's own tool block in either direction
chat template kwargs flag | How llama.cpp passes template options
max_tool_response_chars | Limits how long a tool response can be

Example: You run Qwen3.8-27b at medium effort on repository problems with the Sharp template. On that test the Sharp template needed a median of 20.0 minutes per fix, while the stock one needed 54.6.

For newcomers and experts: If you are new, a chat template is the wrapper that formats your messages for the model. Swapping it changes behavior without touching the weights, so try it on one model and compare.

For experts, this is froggeric's v22.5 plus a spliced system prompt and local revisions. Version 22.4.1 stands down when LM Studio's MLX backend injects its own tool protocol, which the card measured at 6 of 6 tool calls parsed against 2 of 6 before.

You send a chat to the runtime
The Sharp template formats the messages
The model thinks less and answers tighter
Tool calls parse under one protocol

Good to know: The card's SWE-bench-Live test solved 15 of 25 problems with Sharp against 16 with the stock template.

Title: Median minutes to a fix on SWE-bench-Live
Stock template: 54.6 minutes
Sharp template: 20.0 minutes

• License: Apache 2.0, as the Hugging Face record lists it.
• Cost to run: no API key or account is named. You need a local model and runtime.
• Data it reaches: downloading reaches Hugging Face. The template itself only formats messages.
• Not for: models outside Qwen3.5, 3.6 and 3.8. The card also says the SWE-bench-Live gain came from speed, not from more solved problems.

On Hugging Face, 13 November 2026

• Library: mlx
• Downloads in the last 30 days: 0
• Likes: 502
• License: apache-2.0
• Based on: froggeric/Qwen-Fixed-Chat-Templates
• Tags: jinja, chat-template, qwen, qwen3.5, qwen3.6, qwen3.8, llama.cpp, lm-studio, vllm, tool-calling, thinking, token-efficient

Links

• Model: https://huggingface.co/peculiar-ragdoll/Qwen-Sharp-Chat-Templates
• Files: https://huggingface.co/peculiar-ragdoll/Qwen-Sharp-Chat-Templates/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download peculiar-ragdoll/Qwen-Sharp-Chat-Templates

Then follow the usage in its model card. Nothing in this kit runs by itself.