diff --git a/keras2onnx/_builtin.py b/keras2onnx/_builtin.py
index 1cd10c78..0f12a22a 100644
--- a/keras2onnx/_builtin.py
+++ b/keras2onnx/_builtin.py
@@ -20,6 +20,7 @@ class TYPES:
     # tf-node types:
     Identity = 'Identity'
     Const = 'Const'
+    AddN = 'AddN'
     Any = 'Any'
     All = 'All'
     BatchMatMul = 'BatchMatMul'
@@ -170,6 +171,15 @@ def convert_tf_identity(scope, operator, container):
     default_convert(scope, operator, container)
 
 
+@converter_func(TYPES.AddN)
+def convert_tf_addn(scope, operator, container):
+    oopb = OnnxOperatorBuilder(container, scope)
+    oopb.apply_op_with_output("apply_add",
+                              operator.input_full_names,
+                              operator.output_full_names,
+                              name=operator.full_name + '_add')
+
+
 @converter_func(TYPES.BatchToSpaceND)
 def convert_tf_batch_to_space(scope, operator, container):
     node = operator.raw_operator
diff --git a/tests/test_layers.py b/tests/test_layers.py
index 3567ac92..6fc264b9 100644
--- a/tests/test_layers.py
+++ b/tests/test_layers.py
@@ -97,6 +97,21 @@ def test_keras_lambda(self):
         expected = model.predict(data)
         self.assertTrue(run_onnx_runtime('onnx_lambda', onnx_model, data, expected, self.model_files))
 
+    def test_tf_addn(self):
+        input1 = Input(shape=(5, 3, 4), dtype=tf.float32)
+        input2 = Input(shape=(5, 3, 4), dtype=tf.float32)
+        sum = Lambda(tf.add_n)([input1, input2])
+        model = keras.models.Model(inputs=[input1, input2], outputs=sum)
+
+        onnx_model = keras2onnx.convert_keras(model, 'tf_add_n')
+        batch_data1_shape = (2, 5, 3, 4)
+        batch_data2_shape = (2, 5, 3, 4)
+        data1 = np.random.rand(*batch_data1_shape).astype(np.float32)
+        data2 = np.random.rand(*batch_data2_shape).astype(np.float32)
+        expected = model.predict([data1, data2])
+        self.assertTrue(
+            run_onnx_runtime('tf_add_n', onnx_model, [data1, data2], expected, self.model_files))
+
     def test_tf_conv(self):
         model = Sequential()
         k = tf.constant(np.random.normal(loc=0.0, scale=1.0, size=(1, 2, 3, 5)).astype(np.float32))