Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: make Thinking off and Preserve thinking work on Qwen3.6 (#12989)

Nilay02a9b4cf4bca
diff --git a/studio/backend/tests/test_qwen36_own_chat_template.py b/studio/backend/tests/test_qwen36_own_chat_template.py
new file mode 100644
index 00000000000..14caf50b6ba
--- /dev/null
+++ b/studio/backend/tests/test_qwen36_own_chat_template.py
@@ -0,0 +1,95 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parent.parent
+if str(_BACKEND) not in sys.path:
+ sys.path.insert(0, str(_BACKEND))
+
+from utils.datasets.model_mappings import MODEL_TO_TEMPLATE_MAPPER # noqa: E402
+
+MODEL = "unsloth/Qwen3.6-27B"
+
+_QWEN36_STYLE_TEMPLATE = (
+ "{% for m in messages %}{{ '<|im_start|>' + m.role + '\\n' }}"
+ "{% if m.role == 'assistant' and m.reasoning_content and preserve_thinking is defined"
+ " and preserve_thinking is true %}{{ '<think>\\n' + m.reasoning_content + '\\n</think>\\n\\n' }}"
+ "{% endif %}{{ m.content + '<|im_end|>\\n' }}{% endfor %}"
+ "{% if add_generation_prompt %}{{ '<|im_start|>assistant\\n' }}"
+ "{% if enable_thinking is defined and enable_thinking is false %}{{ '<think>\\n\\n</think>\\n\\n' }}"
+ "{% else %}{{ '<think>\\n' }}{% endif %}{% endif %}"
+)
+
+
+def _tokenizer():
+ from tokenizers import Tokenizer, models, pre_tokenizers
+ from transformers import PreTrainedTokenizerFast
+
+ vocab = {"[UNK]": 0, "<|endoftext|>": 1, "<|im_start|>": 2, "<|im_end|>": 3}
+ backend = Tokenizer(models.WordLevel(vocab, unk_token = "[UNK]"))
+ backend.pre_tokenizer = pre_tokenizers.Whitespace()
+ tokenizer = PreTrainedTokenizerFast(
+ tokenizer_object = backend,
+ unk_token = "[UNK]",
+ eos_token = "<|im_end|>",
+ pad_token = "<|endoftext|>",
+ )
+ tokenizer.chat_template = _QWEN36_STYLE_TEMPLATE
+ return tokenizer
+
+
+def _prompt(monkeypatch, messages, **kwargs):
+ try:
+ from core.inference import inference
+ except (ImportError, RuntimeError) as exc:
+ pytest.skip(f"full inference backend unavailable ({type(exc).__name__}: {exc})")
+ # another test on this worker may have bound the module to a stubbed unsloth
+ if getattr(inference.get_chat_template, "__module__", None) != "unsloth.chat_templates":
+ pytest.skip("inference module is bound to a stubbed unsloth.chat_templates")
+ backend = inference.InferenceBackend.__new__(inference.InferenceBackend)
+ backend.active_model_name = MODEL
+ backend.models = {MODEL: {"tokenizer": _tokenizer(), "is_vision": False}}
+ prompts = []
+ monkeypatch.setattr(
+ backend,
+ "generate_stream",
+ lambda prompt, *a, **k: prompts.append(prompt) or iter(()),
+ raising = False,
+ )
+ list(
+ backend._generate_chat_response_inner(
+ messages = messages, system_prompt = "Be brief.", **kwargs
+ )
+ )
+ assert isinstance(prompts[0], str)
+ return prompts[0]
+
+
+def test_thinking_off_reaches_qwen36(monkeypatch):
+ prompt = _prompt(
+ monkeypatch, [{"role": "user", "content": "What is 17 * 23?"}], enable_thinking = False
+ )
+ assert prompt.endswith("<|im_start|>assistant\n<think>\n\n</think>\n\n")
+
+
+def test_preserve_thinking_keeps_earlier_reasoning_for_qwen36(monkeypatch):
+ messages = [
+ {"role": "user", "content": "What is 2 + 2?"},
+ {"role": "assistant", "content": "4", "reasoning_content": "two plus two is four"},
+ {"role": "user", "content": "And times 3?"},
+ ]
+ prompt = _prompt(monkeypatch, messages, enable_thinking = True, preserve_thinking = True)
+ assert "<think>\ntwo plus two is four\n</think>\n\n4<|im_end|>" in prompt
+ assert prompt.endswith("<|im_start|>assistant\n<think>\n")
+
+
+ "name",
+ ["unsloth/Qwen3.6-27B", "Qwen/Qwen3.6-27B", "unsloth/Qwen3.6-35B-A3B", "Qwen/Qwen3.6-35B-A3B"],
+)
+def test_qwen36_keeps_own_template_like_qwen35(name):
+ assert MODEL_TO_TEMPLATE_MAPPER[name.lower()] == MODEL_TO_TEMPLATE_MAPPER["unsloth/qwen3.5-27b"]
diff --git a/studio/backend/utils/datasets/model_mappings.py b/studio/backend/utils/datasets/model_mappings.py
index 3a1c151459a..d93f91c7165 100644
--- a/studio/backend/utils/datasets/model_mappings.py
+++ b/studio/backend/utils/datasets/model_mappings.py
@@ -362,10 +362,6 @@ TEMPLATE_TO_MODEL_MAPPER = {
"unsloth/Qwen3-4B-Thinking-2507-bnb-4bit",
"unsloth/Qwen3-30B-A3B-Thinking-2507",
"Qwen/Qwen3-30B-A3B-Thinking-2507",
- "Qwen/Qwen3.6-35B-A3B",
- "unsloth/Qwen3.6-35B-A3B",
- "Qwen/Qwen3.6-27B",
- "unsloth/Qwen3.6-27B",
),
"qwen3.5": (
"unsloth/Qwen3.5-0.8B",
@@ -374,6 +370,10 @@ TEMPLATE_TO_MODEL_MAPPER = {
"unsloth/Qwen3.5-9B",
"unsloth/Qwen3.5-27B",
"unsloth/Qwen3.5-35B-A3B",
+ "Qwen/Qwen3.6-35B-A3B",
+ "unsloth/Qwen3.6-35B-A3B",
+ "Qwen/Qwen3.6-27B",
+ "unsloth/Qwen3.6-27B",
),
"zephyr": (
"unsloth/zephyr-sft-bnb-4bit",