From eb4e1aa2a4341d4ea8635856f89826a4ac8be93f Mon Sep 17 00:00:00 2001
From: Lynn <lynn_inf@hotmail.com>
Date: Tue, 18 Aug 2026 19:27:31 +0800
Subject: [PATCH] Fix: set enable_thinking to True for qwen3.8-2.4t-a95b
 (#18442)

---
 rag/llm/chat_model.py                         |  5 ++--
 .../llm/test_chat_model_thinking_policy.py    | 27 +++++++++++++++++++
 2 files changed, 30 insertions(+), 2 deletions(-)

diff --git a/rag/llm/chat_model.py b/rag/llm/chat_model.py
index bb50d49c1f..577b9a84a9 100644
--- a/rag/llm/chat_model.py
+++ b/rag/llm/chat_model.py
@@ -152,9 +152,10 @@ def _apply_model_family_policies(
     # Qwen3 keeps RAGFlow's system default of disabling thinking unless explicitly overridden.
     if "qwen3" in model_name_lower:
         _pop_thinking_controls()
-        # -preview variants (e.g. qwen3.8-max-preview) only accept
+        # -preview variants (e.g. qwen3.8-max-preview) and the flagship
+        # reasoning model qwen3.8-2.4t-a95b only accept
         # enable_thinking=True; the API rejects any other value.
-        if "-preview" in model_name_lower:
+        if "-preview" in model_name_lower or "2.4t-a95b" in model_name_lower:
             enable_thinking = True
         else:
             enable_thinking = thinking_type == "enabled" if thinking_type else False
diff --git a/test/unit_test/rag/llm/test_chat_model_thinking_policy.py b/test/unit_test/rag/llm/test_chat_model_thinking_policy.py
index ae512abcee..32ea723905 100644
--- a/test/unit_test/rag/llm/test_chat_model_thinking_policy.py
+++ b/test/unit_test/rag/llm/test_chat_model_thinking_policy.py
@@ -73,6 +73,33 @@ def test_qwen3_preview_ignores_disabled_thinking():
     assert kwargs["extra_body"]["enable_thinking"] is True
 
 
+def test_qwen3_24t_a95b_forces_thinking_true():
+    """qwen3.8-2.4t-a95b (flagship reasoning model) only accepts enable_thinking=True."""
+    gen_conf, kwargs = _apply_model_family_policies(
+        "qwen3.8-2.4t-a95b",
+        backend="base",
+        gen_conf={},
+        request_kwargs={},
+    )
+
+    assert gen_conf == {}
+    assert kwargs["extra_body"]["enable_thinking"] is True
+
+
+def test_qwen3_24t_a95b_ignores_disabled_thinking():
+    """Even with thinking=disabled, qwen3.8-2.4t-a95b still forces enable_thinking=True."""
+    gen_conf, kwargs = _apply_model_family_policies(
+        "qwen3.8-2.4t-a95b",
+        backend="base",
+        gen_conf={"thinking": "disabled", "temperature": 0.2},
+        request_kwargs={},
+    )
+
+    assert "thinking" not in gen_conf
+    assert gen_conf == {"temperature": 0.2}
+    assert kwargs["extra_body"]["enable_thinking"] is True
+
+
 @pytest.mark.parametrize(
     "provider",
     [SupportedLiteLLMProvider.Tongyi_Qianwen, SupportedLiteLLMProvider.Dashscope],