diff --git a/src/google/adk/models/gemma_llm.py b/src/google/adk/models/gemma_llm.py
index a74b7d83743..110b54307af 100644
--- a/src/google/adk/models/gemma_llm.py
+++ b/src/google/adk/models/gemma_llm.py
@@ -35,6 +35,7 @@
 from pydantic import BaseModel
 from pydantic import Field
 from pydantic import ValidationError
+from pydantic_core import to_jsonable_python
 from typing_extensions import override
 
 logger = logging.getLogger('google_adk.' + __name__)
@@ -286,7 +287,7 @@ def _convert_content_parts_for_gemma(
       has_function_response_part = True
       response_text = (
           f'Invoking tool `{func_response.name}` produced:'
-          f' `{json.dumps(func_response.response)}`.'
+          f' `{_function_response_to_json(func_response.response)}`.'
       )
       new_parts.append(Part.from_text(text=response_text))
     elif func_call := part.function_call:
@@ -299,6 +300,34 @@ def _convert_content_parts_for_gemma(
   return new_parts, has_function_response_part, has_function_call_part
 
 
+def _function_response_to_json(response: dict[str, Any] | None) -> str:
+  """Serializes a function response payload as the JSON text Gemma 3 reads.
+
+  The payload holds whatever the tool returned, which can include values the
+  json module cannot encode, such as datetimes, bytes or Pydantic models. They
+  are converted the way a Gemini request carries a function response: Pydantic
+  model fields that are None are dropped (None values in dicts and lists stay
+  null), bytes are base64-encoded, and values Pydantic cannot serialize fall
+  back to their str(). Non-ASCII text is kept as is, so the model reads the
+  characters rather than escape sequences.
+
+  Args:
+    response: The payload of the function response.
+
+  Returns:
+    The payload as a JSON string.
+  """
+  return json.dumps(
+      to_jsonable_python(
+          response,
+          exclude_none=True,
+          bytes_mode='base64',
+          serialize_unknown=True,
+      ),
+      ensure_ascii=False,
+  )
+
+
 def _build_gemma_function_system_instruction(
     function_declarations: list[FunctionDeclaration],
 ) -> str:
diff --git a/tests/unittests/models/test_gemma_llm.py b/tests/unittests/models/test_gemma_llm.py
index 124da4b3120..a0bc8d0eb36 100644
--- a/tests/unittests/models/test_gemma_llm.py
+++ b/tests/unittests/models/test_gemma_llm.py
@@ -12,11 +12,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import datetime
+
 from google.adk import models
+from google.adk.memory.memory_entry import MemoryEntry
 from google.adk.models.gemma_llm import Gemma
 from google.adk.models.google_llm import Gemini
 from google.adk.models.llm_request import LlmRequest
 from google.adk.models.llm_response import LlmResponse
+from google.adk.tools.load_memory_tool import LoadMemoryResponse
 from google.genai import types
 from google.genai.types import Content
 from google.genai.types import Part
@@ -190,6 +194,80 @@ async def test_preprocess_request_with_function_response():
   assert llm_request.contents[0].parts[0].function_call is None
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "response,want_json",
+    [
+        pytest.param(
+            {"city": "München", "forecast": "東京は晴れ"},
+            '{"city": "München", "forecast": "東京は晴れ"}',
+            id="non_ascii_text",
+        ),
+        pytest.param(
+            {"checked_in_at": datetime.datetime(2026, 10, 22, 9, 30)},
+            '{"checked_in_at": "2026-10-22T09:30:00"}',
+            id="datetime",
+        ),
+        pytest.param(
+            {"thumbnail": b"\xff\xd8"},
+            '{"thumbnail": "_9g="}',
+            id="bytes",
+        ),
+        pytest.param(
+            {"cancelled_at": None, "guests": [{"email": None}]},
+            '{"cancelled_at": null, "guests": [{"email": null}]}',
+            id="none_values",
+        ),
+        pytest.param(
+            {
+                "result": LoadMemoryResponse(
+                    memories=[
+                        MemoryEntry(
+                            content=types.Content(
+                                role="user",
+                                parts=[types.Part(text="I live in Paris.")],
+                            ),
+                            author="user",
+                        )
+                    ]
+                )
+            },
+            '{"result": {"memories": [{"content": {"parts": [{"text": "I live'
+            ' in Paris."}], "role": "user"}, "custom_metadata": {}, "author":'
+            ' "user"}]}}',
+            id="pydantic_model",
+        ),
+    ],
+)
+async def test_preprocess_request_with_function_response_converts_values(
+    response, want_json
+):
+  """Tool results reach Gemma as readable JSON, whatever values they hold."""
+  llm_request = LlmRequest(
+      model="gemma-3-1b-it",
+      contents=[
+          types.Content(
+              role="user",
+              parts=[
+                  types.Part(
+                      function_response=types.FunctionResponse(
+                          name="get_booking", response=response
+                      )
+                  )
+              ],
+          )
+      ],
+      config=types.GenerateContentConfig(),
+  )
+
+  await Gemma()._preprocess_request(llm_request)
+
+  assert (
+      llm_request.contents[0].parts[0].text
+      == f"Invoking tool `get_booking` produced: `{want_json}`."
+  )
+
+
 @pytest.mark.asyncio
 async def test_preprocess_request_with_function_call():
   func_call_data = types.FunctionCall(name="get_current_time", args={})