diff --git a/onnx/helper.py b/onnx/helper.py
index 8af88a70b20..21396aa3075 100644
--- a/onnx/helper.py
+++ b/onnx/helper.py
@@ -332,14 +332,22 @@ def make_tensor(
     if data_type == TensorProto.STRING:
         assert not raw, "Can not use raw_data to store string type"
 
+    np_dtype = mapping.TENSOR_TYPE_TO_NP_TYPE[data_type]
+
     # Check number of vals specified equals tensor size
-    storage_type = mapping.TENSOR_TYPE_TO_STORAGE_TENSOR_TYPE[data_type]
-    expected_size = 1 if (not raw) else (mapping.TENSOR_TYPE_TO_NP_TYPE[storage_type].itemsize)
-    # Flatten a numpy array if its rank > 1
+    expected_size = 1
+    if raw:
+        # NumPy doesn't have BFLOAT16. TENSOR_TYPE_TO_NP_TYPE maps it to float32,
+        # which has the wrong itemsize.
+        if data_type == TensorProto.BFLOAT16:
+            expected_size = 2
+        else:
+            expected_size = np_dtype.itemsize
+
     if type(vals) is np.ndarray and len(vals.shape) > 1:
         vals = vals.flatten()
     for d in dims:
-        expected_size = expected_size * d
+        expected_size *= d
 
     if len(vals) != expected_size:
         raise ValueError("Number of values does not match tensor's size. Expected {}, but it is {}. "
@@ -348,14 +356,12 @@ def make_tensor(
     if raw:
         tensor.raw_data = vals
     else:
-        if (data_type == TensorProto.COMPLEX64
-                or data_type == TensorProto.COMPLEX128):
+        if (data_type == TensorProto.COMPLEX64 or data_type == TensorProto.COMPLEX128):
             vals = split_complex_to_pairs(vals)
-        # float16/bfloat16 are stored as uint16
         elif data_type == TensorProto.FLOAT16:
-            vals = np.array(vals).astype(np.float16).view(dtype=np.uint16).flatten().tolist()
+            vals = np.array(vals).astype(np_dtype).view(dtype=np.uint16).flatten().tolist()
         elif data_type == TensorProto.BFLOAT16:
-            vals = list(map(float32_to_bfloat16, np.array(vals).astype(np.float32).flatten().tolist()))
+            vals = list(map(float32_to_bfloat16, np.array(vals).astype(np_dtype).flatten().tolist()))
         field = mapping.STORAGE_TENSOR_TYPE_TO_FIELD[
             mapping.TENSOR_TYPE_TO_STORAGE_TENSOR_TYPE[data_type]]
         getattr(tensor, field).extend(vals)
diff --git a/onnx/mapping.py b/onnx/mapping.py
index f899b0547a8..be5b688fd74 100644
--- a/onnx/mapping.py
+++ b/onnx/mapping.py
@@ -27,7 +27,8 @@
 # Numpy float32 array is only reversed to TensorProto.FLOAT
 NP_TYPE_TO_TENSOR_TYPE = {v: k for k, v in TENSOR_TYPE_TO_NP_TYPE.items() if k != TensorProto.BFLOAT16}
 
-# This map indicates what storage-type is used in the protobuf (serialized) representation for TensorProto
+# This is only used to get keys into STORAGE_TENSOR_TYPE_TO_FIELD.
+# TODO(https://github.com/onnx/onnx/issues/4261): Remove this.
 TENSOR_TYPE_TO_STORAGE_TENSOR_TYPE = {
     int(TensorProto.FLOAT): int(TensorProto.FLOAT),
     int(TensorProto.UINT8): int(TensorProto.INT32),
diff --git a/onnx/test/helper_test.py b/onnx/test/helper_test.py
index 3c8b88540b1..a0fcacad36f 100644
--- a/onnx/test/helper_test.py
+++ b/onnx/test/helper_test.py
@@ -392,6 +392,28 @@ def test_make_tensor(self) -> None:
         )
         self.assertEqual(string_list, list(tensor.string_data))
 
+    def test_make_int8_tensor(self) -> None:
+        np_array = np.random.randn(2, 3).astype(np.int8)
+
+        tensor = helper.make_tensor(
+            name='test',
+            data_type=TensorProto.INT8,
+            dims=(2, 3),
+            vals=np_array
+        )
+        self.assertEqual(tensor.name, 'test')
+        np.testing.assert_equal(np_array, numpy_helper.to_array(tensor))
+
+        # use raw_data field to store the data
+        tensor = helper.make_tensor(
+            name='test',
+            data_type=TensorProto.INT8,
+            dims=(2, 3),
+            vals=np_array.tobytes(),
+            raw=True,
+        )
+        np.testing.assert_equal(np_array, numpy_helper.to_array(tensor))
+
     def test_make_float16_tensor(self) -> None:
         np_array = np.random.randn(2, 3).astype(np.float16)
 
@@ -404,7 +426,7 @@ def test_make_float16_tensor(self) -> None:
         self.assertEqual(tensor.name, 'test')
         np.testing.assert_equal(np_array, numpy_helper.to_array(tensor))
 
-    def test_make_float16_tensor_with_raw(self) -> None:
+    def test_make_float16_tensor_raw(self) -> None:
         np_array = np.random.randn(2, 3).astype(np.float16)
 
         tensor = helper.make_tensor(
@@ -443,7 +465,7 @@ def test_make_bfloat16_tensor(self) -> None:
         self.assertEqual(tensor.name, 'test')
         np.testing.assert_equal(np_results, numpy_helper.to_array(tensor))
 
-    def test_make_bfloat16_tensor_with_raw(self) -> None:
+    def test_make_bfloat16_tensor_raw(self) -> None:
         # numpy doesn't support bf16, so we have to compute the correct result manually
         np_array = np.array([[1.0, 2.0], [3.0, 4.0], [0.099853515625, 0.099365234375], [0.0998535081744, 0.1], [np.nan, np.inf]],
             dtype=np.float32)