Qwen-Fixed-Chat-Templates

Qwen-Fixed-Chat-Templates

Swaps Qwen's official chat templates for one drop-in Jinja file that fixes rendering errors, cache invalidation, wasted tokens and agent stalls. For people who run Qwen 3.5, 3.6 and 3.8 models locally.

Made by froggeric 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: This is a single Jinja chat template, chat_template.jinja, for Qwen 3.5, 3.6 and 3.8 models, including the 3.8 Flash-Next sizes. The author says the official templates break in several ways, such as rendering errors, a KV cache that gets thrown away, wasted tokens, empty think blocks and agents that stall. This version is v22.5.

Who it suits: It suits people who serve Qwen models on their own machines and connect coding agents or chat apps to them.

Industry | Who uses it | How they use it | Example
Local AI hobby | Enthusiasts | Replace the template in LM Studio for cleaner chats | Pasting the file into the prompt template box
Software development | Developers | Run coding agents against a local Qwen server | Connecting OpenCode or Claude Code to llama-server
Platform engineering | Inference engineers | Serve Qwen models with vLLM or SGLang | Swapping the chat template string in tokenizer_config.json
Model tooling | Template maintainers | Check which template version a model folder uses | Running the bundled check script

You can skip it if your engine cannot load Hugging Face Jinja templates or you do not use Qwen models.

Where to use it:

• LM Studio
• llama.cpp, llama-server and KoboldCPP
• vLLM and SGLang
• MLX and oMLX
• Any other engine that reads Hugging Face Jinja templates

Use case | Who | What they get back
Agent tool calls | Agent users | Tool arguments that render without crashes in history
Reasoning control | API users | A default of medium effort, with switches for more or none
Cache reuse | Server operators | Prompts that keep the KV cache valid between turns

How to set it up: You need a Qwen model and one of the supported engines. The method depends on the engine you use.

• A Qwen 3.5, 3.6 or 3.8 model
• One supported engine such as llama.cpp, LM Studio or vLLM
• The chat_template.jinja file from this repository

1. Download the repository files.
2. For llama.cpp, start llama-server with the template file and the DeepSeek reasoning format.
3. For LM Studio, open the model, find Prompt Template, replace its contents with the file and save.
4. For vLLM, replace the chat_template string in tokenizer_config.json with the one-line version, then serve the model with the parsers shown.

pip install -U huggingface_hub
hf download froggeric/Qwen-Fixed-Chat-Templates

llama-server -m your_model.gguf --jinja --chat-template-file chat_template.jinja --reasoning-format deepseek

vllm serve Qwen/Qwen3.8-2.4T-A95B --reasoning-parser qwen3 --tool-call-parser qwen3_xml

How to use it: Once the template is loaded, you chat as normal. Reasoning depth defaults to medium. You can switch it off with enable_thinking set to false or reasoning_effort set to none, or steer it with inline tags typed into the chat.

Tool | What it does
--chat-template-file | Loads this template into llama-server
--reasoning-format deepseek | Moves think blocks into the reasoning_content field of the response
--reasoning-preserve | Keeps the prefix KV cache on recent llama.cpp builds
scripts/check_applied.py | Inspects model folders and GGUF files for the active template version

Example: You connect a coding agent such as Claude Code to llama-server using the DeepSeek reasoning format. The thinking text goes to its own response field, so it no longer leaks into the text stream and cuts off tool calls.

For newcomers and experts: If you are new to this, a chat template is the recipe that turns your messages into the exact text a model reads. A broken recipe causes odd errors, so a first small try is loading this file in LM Studio.

For people who know the field, the template maps client effort names onto Qwen's levels, so high, max, ultracode and extreme become xhigh, and none or off disables thinking. It also merges consecutive leading system and developer messages and applies two-tier tool error recovery.

Your engine loads chat_template.jinja
Messages and tools are rendered in one format
Reasoning effort is set by default or by a tag
The model replies and thinking is split out

Good to know: The official Qwen 3.8 template defaults to the highest effort, while this one defaults to medium to save tokens and keep the cache stable.

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: only the prompts your engine formats, plus model folders the check script reads
• Not for: engines that cannot load Hugging Face Jinja templates

On Hugging Face, 13 November 2026

• Library: mlx
• Downloads in the last 30 days: 0
• Likes: 1,901
• License: apache-2.0
• Fine-tunes: 12
• Tags: jinja, chat-template, qwen, qwen3.5, qwen3.6, qwen3.8, lm-studio, llama.cpp, vllm, tool-calling, thinking

Links

• Model: https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates
• Files: https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download froggeric/Qwen-Fixed-Chat-Templates

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