--- license: apache-2.0 tags: - jinja - chat-template - qwen - qwen3.5 - qwen3.6 - lm-studio - mlx - llama.cpp - tool-calling - thinking --- # Fixed Chat Templates for Qwen 3.5 & 3.6 > **2026-05-05** — Reviewed against community merged templates (allanchan339, fakezeta). Confirmed all useful features already present; `from_json` string-arg parsing not portable to C++ engines. Added auto-close unclosed thinking, thanks to allanchan339. Drop-in Jinja templates that fix rendering errors, token waste, and missing features in the official Qwen chat templates. Works in LM Studio, llama.cpp, vLLM, MLX, oMLX, and any engine that supports HuggingFace Jinja templates. ## Why you need this The official Qwen templates have bugs that break real usage: | Problem | Impact | |---------|--------| | Tool calls fail on C++ engines | `|items` filter doesn't exist in LM Studio, llama.cpp, MLX, oMLX — tool calls produce a template error | | `developer` role rejected | Modern APIs send it; the official template raises an error | | Empty thinking blocks spam context | Every past turn gets wrapped in tags, even with nothing inside | | No way to toggle thinking | You're stuck with whatever the model defaults to | | Qwen 3.6: `` hallucination | Model sometimes generates the wrong closing tag; parser fails | | No-user-query exception breaks tool calling | `raise_exception` crashes agentic loops and resets in OpenClaw and similar runtimes | | Unclosed thinking before tool call | Model starts reasoning then calls a tool without closing thinking block — malformed output | All seven are fixed here, plus a clean `<|think_on|>` / `<|think_off|>` toggle you can drop into any message. ## Quick install ### LM Studio 1. Open your Qwen model in the right-side panel 2. Scroll to **Prompt Template** 3. Replace the template with the contents of `qwen3.5/chat_template.jinja` or `qwen3.6/chat_template.jinja` 4. Save ### llama.cpp / koboldcpp ```bash --jinja --chat-template-file qwen3.6/chat_template.jinja ``` ### vLLM / TextGen Replace the `chat_template` string in your `tokenizer_config.json` with the file contents. ### oMLX Overwrite `chat_template.jinja` in your local model directory. Load with `--jinja`. Remove any `chat_template_kwargs` overrides — the template handles everything internally. ## Which file do I use? | File | For models | |------|-----------| | [`qwen3.5/chat_template.jinja`](qwen3.5/chat_template.jinja) | Qwen3.5-35B-A3B, Qwen3.5-32B, Qwen3.5-14B, and all Qwen 3.5 variants | | [`qwen3.6/chat_template.jinja`](qwen3.6/chat_template.jinja) | Qwen3.6-27B, Qwen3.6-35B-A3B, and all Qwen 3.6 variants | The 3.6 template is a superset — it additionally handles `preserve_thinking`, `` hallucination recovery, and interrupted thought streams. If you're on 3.6, use the 3.6 file. ## Thinking toggle Drop `<|think_on|>` or `<|think_off|>` anywhere in your system or user prompt. The template intercepts the tag, removes it from context so the model never sees it, and flips the mode. **Fast answer, no reasoning:** ``` System: You are a coding assistant. <|think_off|> User: What's 2+2? ``` **Deep reasoning:** ``` System: You are a coding assistant. <|think_on|> User: Implement a red-black tree in Rust. ``` The tag syntax (`<|think_on|>`, `<|think_off|>`) uses Qwen's control-token delimiters, so it will never collide with real text. Earlier community templates used `/think`, which broke legitimate paths like `cd /mnt/project/think`. ## Pre-installed models These templates are already bundled with: - [froggeric/Qwen3.6-27B-MLX-8bit](https://huggingface.co/froggeric/Qwen3.6-27B-MLX-8bit) - [froggeric/Qwen3.6-27B-MLX-4bit](https://huggingface.co/froggeric/Qwen3.6-27B-MLX-4bit) - [froggeric/Qwen3.5-35B-A3B-Uncensored-FernflowerAI-MLX-8bit](https://huggingface.co/froggeric/Qwen3.5-35B-A3B-Uncensored-FernflowerAI-MLX-8bit) - [froggeric/Qwen3.5-35B-A3B-Uncensored-FernflowerAI-MLX-4bit](https://huggingface.co/froggeric/Qwen3.5-35B-A3B-Uncensored-FernflowerAI-MLX-4bit) - [froggeric/Qwen3.6-27B-Uncensored-Heretic-v2-MLX-8bit](https://huggingface.co/froggeric/Qwen3.6-27B-Uncensored-Heretic-v2-MLX-8bit) - [froggeric/Qwen3.6-27B-Uncensored-Heretic-v2-MLX-6bit](https://huggingface.co/froggeric/Qwen3.6-27B-Uncensored-Heretic-v2-MLX-6bit) - [froggeric/Qwen3.6-27B-Uncensored-Heretic-v2-MLX-4bit](https://huggingface.co/froggeric/Qwen3.6-27B-Uncensored-Heretic-v2-MLX-4bit) - [froggeric/Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-8bit](https://huggingface.co/froggeric/Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-8bit) - [froggeric/Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-6bit](https://huggingface.co/froggeric/Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-6bit) - [froggeric/Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit](https://huggingface.co/froggeric/Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit) If you're using one of those, you already have the template. This repo is for everyone else. ---
Technical details — what exactly was fixed ## Tool calls on C++ engines The official template iterates tool call arguments with `|items`: ```jinja {%- for key, value in tool_call.arguments|items %} ``` Python's Jinja supports `|items`. C++ runtimes (LM Studio, llama.cpp, MLX) do not — the template produces a rendering error instead of output. This template uses direct dictionary key lookups instead: ```jinja {%- for args_name in tool_call.arguments %} {%- set args_value = tool_call.arguments[args_name] %} ``` It also replaces `is sequence` with `is iterable` (stricter C++ runtimes require it), removes `|safe` wrappers (also Python-only), and handles arguments returned as raw strings instead of objects. ## `developer` role The OpenAI-compatible API spec sends `message.role == "developer"` for system-level instructions. The official Qwen template only checks for `"system"` and throws on anything else. Both templates here accept `"developer"` and map it to the system role. ## Empty thinking blocks The official template wraps every past assistant turn in thinking tags: ``` <|im_start|>assistant Here is the answer... ``` When there's no reasoning content, those tags are dead weight — they waste context tokens and break prefix caching. The Qwen 3.5 template checks `reasoning_content` before emitting. The Qwen 3.6 template goes further: it respects the `preserve_thinking` kwarg, checks `reasoning_content|trim|length > 0`, and ties history visibility to the `<|think_off|>` override. ## `` hallucination (Qwen 3.6 only) The Qwen 3.6 model sometimes generates `` instead of the expected closing tag. The official parser splits on `` only and fails. The 3.6 template detects which closing tag was actually used and splits on that: ```jinja {%- if '' in content %} {%- set think_end_token = '' %} {%- elif '' in content %} {%- set think_end_token = '' %} ``` It also handles interrupted generation (max tokens hit mid-thought) by rescuing incomplete streams instead of injecting broken tag pairs. ## Arguments serialization The official template serializes argument values with `|tojson` unconditionally, which turns Python `True` into JSON `true` correctly but fails when the value is already a string. The fixed templates check the type first — strings pass through as-is, everything else goes through `|tojson`. ## Auto-close unclosed thinking before tool calls The model sometimes starts a thinking block and then immediately calls a tool without emitting the closing tag. The official template doesn't handle this — the unclosed thinking tag bleeds into the tool call, producing malformed output. Both fixed templates detect this pattern and auto-inject the closing tag before the tool call boundary. ## No-user-query exception The official template scans the message list in reverse to find the last "real" user query (skipping tool-result wrappers). If all user messages are tool results — or there are no user messages at all — it fires `raise_exception('No user query found in messages.')` and the template **hard-crashes**. This breaks real usage: - **Agentic tool-calling chains** where the conversation ends with tool results and no fresh user query - **After `/reset` or `/new`** in runtimes like OpenClaw, where tool results from a prior session persist without a new user message - **System-only contexts** with no user messages The fix replaces the exception with a graceful fallback: `{%- set ns.last_query_index = messages|length - 1 %}`. The thinking display logic then degrades naturally — assistant turns with reasoning content still show thinking tags when `preserve_thinking` is enabled.
Comparison — Qwen 3.5 templates | Feature | Official | LuffyTheFox | mod-ellary | Pneuny | **This** | |---------|----------|-------------|------------|--------|----------| | Tool arguments | Fails | Fixed | Missing | Fixed | **Fixed** | | `\|safe` removed | Fails | Fixed | Missing | Fixed | **Fixed** | | `developer` role | Missing | Missing | Missing | Missing | **Added** | | Thinking toggle | None | None | `/think` (system only) | None | **`<\|think_off\|>` anywhere** | | Empty think in history | Broken | Broken | Tags omitted | Broken | **Fixed** | | Text safety | N/A | N/A | Breaks on `/think` in paths | N/A | **Safe** | | Clean instructions | Yes | Yes | Yes | Injects "I cannot call a tool" | **Yes** | | No-user-query crash | Crashes | Crashes | Crashes | Crashes | **Graceful fallback** | | Auto-close thinking before tool | Not handled | Not handled | Not handled | Not handled | **Auto-injects close tag** |
Comparison — Qwen 3.6 template | Feature | Official | **This** | |---------|----------|----------| | Tool arguments | Fails (`\|items`) | **Fixed** | | `\|safe` removed | Fails | **Fixed** | | `developer` role | Missing | **Added** | | Thinking toggle | None | **`<\|think_off\|>` anywhere** | | `preserve_thinking` | Spams empty blocks | **Dynamic length checks** | | `` hallucination | Fails | **Detected and handled** | | Interrupted streams | Broken tags | **Rescued** | | Auto-close thinking before tool | Not handled | **Auto-injects close tag** | | No-user-query crash | Crashes | **Graceful fallback** |
--- ## Authorship | Role | Author | |------|--------| | Original models | Alibaba Cloud (Qwen team) | | Template fixes | [froggeric](https://huggingface.co/froggeric) | ## License Apache-2.0, inherited from Qwen.