diff --git a/fastembed/common/preprocessor_utils.py b/fastembed/common/preprocessor_utils.py
index fcd8acaa..fb853924 100644
--- a/fastembed/common/preprocessor_utils.py
+++ b/fastembed/common/preprocessor_utils.py
@@ -113,8 +113,10 @@ def load_tokenizer(model_dir: Path) -> tuple[Tokenizer, dict[str, int]]:
 
     # Padding is always normalized to batch-longest. A serialized fixed length shorter than the
     # truncation limit leaves longer encodings untouched, which produces ragged batches, and a
-    # fixed length equal to it pads every batch to the maximum. Direction and pad token metadata
-    # are taken from the serialized settings, since some models pad on the left.
+    # fixed length equal to it pads every batch to the maximum. Pad token metadata is taken from
+    # the serialized settings. The direction follows transformers: `padding_side` from
+    # tokenizer_config.json wins over the serialized direction, since some models pad on the left
+    # and set it only in the config.
     padding = tokenizer.padding or {}
     pad_token = padding.get("pad_token") or tokenizer_config.get("pad_token")
     if pad_token is None:
@@ -129,7 +131,7 @@ def load_tokenizer(model_dir: Path) -> tuple[Tokenizer, dict[str, int]]:
         raise ValueError(f"Could not resolve an id for the pad token {pad_token!r} in {model_dir}")
 
     tokenizer.enable_padding(
-        direction=padding.get("direction", "right"),
+        direction=tokenizer_config.get("padding_side") or padding.get("direction", "right"),
         pad_id=pad_id,
         pad_type_id=padding.get("pad_type_id", 0),
         pad_token=pad_token,
diff --git a/tests/test_preprocessor_utils.py b/tests/test_preprocessor_utils.py
index b9f718a8..d04f0312 100644
--- a/tests/test_preprocessor_utils.py
+++ b/tests/test_preprocessor_utils.py
@@ -236,6 +236,30 @@ def test_truncation_direction_resolution(
     assert tokenizer.truncation["direction"] == expected
 
 
+@pytest.mark.parametrize(
+    "serialized_direction,padding_side,expected",
+    [
+        (None, None, "right"),
+        ("left", None, "left"),
+        (None, "left", "left"),
+        ("left", "right", "right"),  # tokenizer_config.json wins, as in transformers
+    ],
+)
+def test_padding_direction_resolution(
+    make_model_dir, serialized_direction, padding_side, expected
+) -> None:
+    model_dir = make_model_dir(
+        tokenizer_config={"padding_side": padding_side},
+        padding=None
+        if serialized_direction is None
+        else {"pad_id": 0, "pad_token": "[PAD]", "direction": serialized_direction},
+    )
+
+    tokenizer, _ = load_tokenizer(model_dir)
+
+    assert tokenizer.padding["direction"] == expected
+
+
 @pytest.mark.parametrize(
     "model_max_length,max_length",
     [