The Underrated Lifesaving Template for Qwen Local Deployment

When trying to run Qwen 3.6 for agent tasks, the author repeatedly hit parsing errors in LM Studio, llama.cpp and vLLM, discovering that the root cause was the official Qwen Jinja chat template.

What the community fixed

Froggeric forked the official template into Qwen‑Fixed‑Chat‑Templates , now at version v19, which works as a drop‑in replacement for the entire Qwen 3.5/3.6 series (27B/32B/35B). The new template eliminates rendering bugs, KV‑Cache invalidation, token waste and fatal agent dead‑locks.

Supported inference engines

LM Studio : replace the Prompt Template in the right‑hand panel with the new chat_template.jinja and save.

llama.cpp / koboldcpp : launch with --jinja --chat-template-file chat_template.jinja.

vLLM : replace the chat_template field inside tokenizer_config.json and add --tool-call-parser qwen3_coder.

MLX / oMLX : overwrite the local chat_template.jinja and start with --jinja, removing any chat_template_kwargs overrides.

Any engine that supports HuggingFace Jinja templates works as well.

Five categories of bugs in the official template

Agent dead‑loop : premature stop, retry spin, over‑thinking after tool calls, and mis‑interpreting any error token as a tool failure.

Performance : KV‑Cache loss due to history pruning each round, and the “empty <think> ” poisoning that makes the model think it can call tools without thinking.

Compatibility : crashes on old C++ engines (e.g., loop.previtem), mismatched tool‑call XML vs. JSON formats, and Jinja C++ crashes caused by Python‑only filters like map and first.

Stability : crashes when inserting system messages, when no user message is present, or when stray <think> blocks leak into the tool parser.

Edge cases : the official template rejects the developer role used by Claude Code/Codex/OpenCode, ignores the --reasoning off flag, and sometimes hallucates extra reasoning tags.

What v19 changes (The Agentic Loop Cure)

Remove empty <think> poisoning : the previous shortcut cleared the block to an empty tag, causing >80% of premature‑stop bugs; v19 rewrites the AST to never inject empty think blocks.

Eliminate the system‑prompt logical trap : the old <IMPORTANT> forced a mandatory </think> before tool calls, making the model panic in pure chat; v19 replaces it with a Universal Synthesis instruction that allows direct replies after </think>.

KV‑Cache 100 % hit + Amnesia fix : the new default preserve_thinking=true keeps the thinking chain in order, fully curing multi‑step tool loops’ “memory loss” and guaranteeing 100 % prefix KV‑Cache hit rate, which speeds up local inference.

These fixes let the agent loop run to completion without unexpected detours.

Installation guide per engine

LM Studio

1. Open the Qwen model in the right panel
2. Locate Prompt Template
3. Paste the entire chat_template.jinja
4. Save

llama.cpp / koboldcpp

--jinja --chat-template-file chat_template.jinja

vLLM

# Replace tokenizer_config.json's chat_template field with the Jinja file
--tool-call-parser qwen3_coder

oMLX

Overwrite chat_template.jinja in the model directory
Start with --jinja (remove any chat_template_kwargs overrides)

Thought‑mode toggle

The template supports on‑the‑fly switching of the thinking mode by inserting control tokens in system or user messages:

System: You are a coding assistant. <|think_off|>
User: What's 2+2?

System: You are a coding assistant. <|think_on|>
User: Implement a red‑black tree in Rust.

The delimiter <|think_on|> never collides with normal text or file paths, offering a higher safety level.

Saving tokens

v19 enables preserve_thinking=true by default for maximum KV‑Cache hit. If GPU memory or context window is tight, set it to false in the engine kwargs:

{
  "preserve_thinking": false
}

Disabling it reduces KV‑Cache hit rate for multi‑turn dialogs but saves context space.

Conclusion

The project is a straightforward, drop‑in fix that patches every known hole in the official template. It is ideal for anyone running Qwen 3.5/3.6 locally with llama.cpp, LM Studio, vLLM or MLX, especially developers needing reliable agent/tool‑call behavior and stable KV‑Cache performance. The only friction point is the manual template replacement, which the author notes can be done in about five minutes following the official README.