license	language	base_model	pipeline_tag	tags
apache-2.0	en	Qwen/Qwen3.5-35B-A3B	text-generation	qwen chat-template jinja qwen3.5 llama-cpp open-webui vllm tool-calling streaming

Qwen 3.5 Jinja Chat Template v1 attuned by Barubary

A Jinja chat template for all Qwen 3.5 models on llama.cpp, Open WebUI, vLLM, Ollama, LM Studio, and any OAI-compatible endpoint.

Created this because of some current projects related to this model.

21 fixes over the official Qwen 3.5 chat template — addressing bugs that are still open upstream as of March 2026.

Active Bug Reports Fixed

This template directly addresses the following community-reported bugs:

Bug Report	Platform	Fix
Tool calling chat template is broken	HuggingFace	Fix 6
Parallel tool calls interleaving	GitHub	Fix 15
KV-cache reuse breaks with enable_thinking=false	GitHub	Fix 12
Cannot close thinking via enable_thinking: false	GitHub	Fix 1, 19
Missing reasoning_content in Tool Calling	GitHub	Fix 13
LM Studio parser breaks Qwen3.5 tool calling	Reddit	Fix 18, 19
Qwen3.5 27B getting stuck in loops	Reddit	Fix 17
Template problem	HuggingFace	Fix 6, 7

All 21 Fixes

Each fix is labeled inline in the template source (e.g., {#- FIX6 #}).

#	Fix	What It Solves
1	add_vision_id / enable_thinking safe defaults	Crashes when config vars not passed
2	Precomputed _last_idx for namespace() constructor	llama.cpp minja parser compatibility
3	Developer role handled	Claude Code / Codex / OpenCode support
4	System/developer split before main loop	Duplicate system messages
5	item.type checked before 'in item' key test	Type-check ordering bug
6	arguments.items() replaces bare key iteration	Tool calling crash (HF discussion #4)
7	| safe filter removed	llama.cpp compatibility
8	tojson/string explicit if/else	No chained filters, prevents double-escaping
9	String arguments pass-through	OAI-compatible proxy support
10	tc alias avoids shadowing tool_call loop var	Variable scoping bug
11	ns2 namespace replaces loop.previtem / loop.nextitem	llama.cpp minja doesn't support loop helpers
12	enable_thinking applied to in-context assistant turns	KV-cache reuse bug (GitHub #1826)
13	reasoning_content is defined + not none guard	Missing reasoning_content (GitHub #26)
14	loop.index0 > (not >=) for assistant thinking scope	Off-by-one in thinking block placement
15	Parallel tool calls: \n\n delimiter between blocks	Parallel tool call interleaving (GitHub #7117)
16	Long tool args/responses: configurable truncation guard	Context overflow from massive tool outputs
17	Deep agent loops: graceful fallback to index 0	Agent loops crashing after 5+ hops
18	Streaming compat: clean newline boundaries on all XML tags	LM Studio parser breaks (Reddit)
19	Auto-disable thinking when tools active	<tool_call> leaks into <think> blocks
20	Unknown roles: graceful fallback mapped to user role	Planner/critic/custom roles crash
21	Flattened nesting depth; _has_tools precomputed	llama.cpp minja stability

Feature Comparison

Feature	This Template	Official Qwen	Unsloth	bartowski
Parallel tool call separation	✅	❌	❌	❌
Auto-disable thinking with tools	✅	❌	❌	❌
Deep agent loop fallback	✅	❌	❌	❌
Unknown role graceful fallback	✅	❌	❌	❌
Configurable truncation guards	✅	❌	❌	❌
Streaming-safe XML boundaries	✅	❌	❌	partial
Developer role support	✅	❌	❌	❌
arguments.items() fix	✅	❌	✅	❌
reasoning_content guard	✅	❌	partial	❌

Usage

llama-server (llama.cpp)

llama-server \
  -m Qwen3.5-35B-A3B-*.gguf \
  --jinja -fa \
  --chat-template-file chat_template.jinja \
  -c 32768 -ngl 99 \
  --temp 0.6 --top-k 20 --top-p 0.8 \
  --cache-type-k q8_0 --cache-type-v q8_0 \
  --host 0.0.0.0 --port 8080

Open WebUI

Mount the template via Docker:

volumes:
  - ./chat_template.jinja:/templates/chat_template.jinja:ro
command: >
  --chat-template-file /templates/chat_template.jinja

vLLM

vllm serve Qwen/Qwen3.5-35B-A3B \
  --chat-template ./chat_template.jinja

Ollama

Copy chat_template.jinja into your Modelfile or use with a compatible frontend.

Configuration

Pass via --chat-template-kwargs:

{
  "enable_thinking": true,
  "auto_disable_thinking_with_tools": true,
  "add_vision_id": false,
  "max_tool_arg_chars": 0,
  "max_tool_response_chars": 8192
}

Variable	Default	Description
enable_thinking	true	Controls <think> mode
auto_disable_thinking_with_tools	true	Auto-disables thinking when tools are provided to prevent <tool_call> bleed into <think> blocks
add_vision_id	false	Prefix images/videos with "Picture N:" / "Video N:"
max_tool_arg_chars	0 (unlimited)	Truncate tool arguments beyond this length
max_tool_response_chars	0 (unlimited)	Truncate tool responses beyond this length

Before / After

Tool Call Bleed Bug (Fix 19)

Before (official template):

<think>
The user wants me to search for...
<tool_call>          ← WRONG: tool call inside think block
<function=search>

After (this template):

<think>

</think>

<tool_call>          ← Correct: thinking auto-disabled when tools present
<function=search>

Parallel Tool Calls (Fix 15)

Before (official template):

<tool_call><function=multiply>...</function></tool_call><tool_call><function=add>...</function></tool_call>

After (this template):

<tool_call>
<function=multiply>
...
</function>
</tool_call>

<tool_call>
<function=add>
...
</function>
</tool_call>

Compatible Models

Tested and compatible with all Qwen 3.5 models:

• Qwen3.5-35B-A3B (all quants)
• Qwen3.5-27B-A3B
• Qwen3.5-14B-A3B
• Qwen3.5-9B
• Qwen3.5-4B
• Qwen3.5-Coder series

Also backward-compatible with Qwen3 32B.

Tested Platforms

• ✅ llama.cpp (b4242+)
• ✅ Open WebUI (v0.4.8+)
• ✅ vLLM (v0.6.4+)
• ✅ Ollama (v0.5.0+)
• ✅ LM Studio (v0.3.5+)
• ✅ Text Generation WebUI (oobabooga)

Credits

Base template architecture: Qwen team (official Qwen3.5 chat template)

All 21 fixes: Barubary (original implementations)

License

Apache 2.0 (same as the official Qwen3.5 template)

Contributing

Found a bug? Open an issue with:

• Minimal reproduction case
• Error logs
• Model and runtime versions

Pull requests welcome.