Send thinking level for Gemma 4 models in Google Generative AI (#184275)

Co-authored-by: Franck Nijhof <git@frenck.dev>

homeassistant/components/google_generative_ai_conversation/entity.py

@@ -110,6 +110,12 @@ def _is_gemini_3_model(model: str) -> bool:
     return name.startswith("gemini-3")
 
 
+def _is_gemma_4_model(model: str) -> bool:
+    """Check if the model is a Gemma 4 series model."""
+    name = model.removeprefix("models/")
+    return name.startswith("gemma-4")
+
+
 def _create_thinking_config(
     model: str,
     thinking_budget: int,
@@ -123,10 +129,21 @@ def _create_thinking_config(
             -1 = automatic (default behavior),
             0 = disable thinking,
             >0 = custom token budget (Gemini 2.5 only).
-        thinking_level: The user-configured thinking level for Gemini 3 models:
-            "auto" = automatic (default), "minimal", "low", "medium", "high".
+        thinking_level: The user-configured thinking level for Gemini 3 and
+            Gemma 4 models: "auto" = automatic (default), "minimal", "low",
+            "medium", "high".
 
     """
+    if _is_gemma_4_model(model):
+        # Gemma 4 only supports the minimal and high thinking levels
+        gemma_level_map: dict[str, ThinkingLevel] = {
+            "minimal": ThinkingLevel.MINIMAL,
+            "high": ThinkingLevel.HIGH,
+        }
+        if thinking_level and thinking_level in gemma_level_map:
+            return ThinkingConfig(thinking_level=gemma_level_map[thinking_level])
+        return None
+
     if not _is_thinking_model(model):
         return None
 

tests/components/google_generative_ai_conversation/test_conversation.py

@@ -4,7 +4,7 @@ import datetime
 from unittest.mock import AsyncMock, patch
 
 from freezegun import freeze_time
-from google.genai.types import GenerateContentResponse, ThinkingLevel
+from google.genai.types import GenerateContentResponse, ThinkingConfig, ThinkingLevel
 import pytest
 from syrupy.assertion import SnapshotAssertion
 
@@ -881,6 +881,37 @@ def test_create_thinking_config_gemini3_auto(
     assert result.thinking_level is None
 
 
+@pytest.mark.parametrize(
+    ("model", "thinking_level", "expected"),
+    [
+        pytest.param(
+            "models/gemma-4-26b-a4b-it",
+            "minimal",
+            ThinkingConfig(thinking_level=ThinkingLevel.MINIMAL),
+            id="minimal",
+        ),
+        pytest.param(
+            "gemma-4-31b-it",
+            "high",
+            ThinkingConfig(thinking_level=ThinkingLevel.HIGH),
+            id="high",
+        ),
+        # The API rejects any other thinking level for Gemma 4
+        pytest.param("models/gemma-4-31b-it", "low", None, id="low"),
+        pytest.param("models/gemma-4-31b-it", "medium", None, id="medium"),
+        pytest.param("models/gemma-4-31b-it", "auto", None, id="auto"),
+        pytest.param("models/gemma-4-31b-it", None, None, id="unset"),
+    ],
+)
+def test_create_thinking_config_gemma4(
+    model: str,
+    thinking_level: str | None,
+    expected: ThinkingConfig | None,
+) -> None:
+    """Test Gemma 4 models only send the supported thinking levels."""
+    assert _create_thinking_config(model, 0, thinking_level) == expected
+
+
 @pytest.mark.parametrize(
     ("model", "thinking_budget", "expected_budget"),
     [