diff --git a/pytorch_forecasting/models/samformer/_samformer_v2.py b/pytorch_forecasting/models/samformer/_samformer_v2.py
index 7d859921e..ee0fdc70c 100644
--- a/pytorch_forecasting/models/samformer/_samformer_v2.py
+++ b/pytorch_forecasting/models/samformer/_samformer_v2.py
@@ -30,6 +30,12 @@ class Samformer(BaseModel):
         Whether to use Reverse Instance Normalization. Default is True.
     persistence_weight : float, optional
         Weight for persistence baseline. Default is 0.0.
+    metadata : dict
+        Dataset metadata produced by
+        :class:`~pytorch_forecasting.data.data_module\
+.EncoderDecoderTimeSeriesDataModule`.
+        Must contain ``"max_encoder_length"``, ``"max_prediction_length"``,
+        and ``"encoder_cont"``.
     """
 
     @classmethod
@@ -58,6 +64,9 @@ def __init__(
         metadata: dict | None = None,
         **kwargs,
     ):
+        if metadata is None:
+            raise ValueError("metadata is required")
+
         super().__init__(
             loss=loss,
             logging_metrics=logging_metrics,
diff --git a/tests/test_models/test_samformer_v2.py b/tests/test_models/test_samformer_v2.py
new file mode 100644
index 000000000..354af0d51
--- /dev/null
+++ b/tests/test_models/test_samformer_v2.py
@@ -0,0 +1,10 @@
+import pytest
+import torch.nn as nn
+
+from pytorch_forecasting.models.samformer._samformer_v2 import Samformer
+
+
+def test_samformer_requires_metadata():
+    """Test that Samformer rejects missing metadata explicitly."""
+    with pytest.raises(ValueError, match="metadata is required"):
+        Samformer(loss=nn.MSELoss(), hidden_size=512, use_revin=True, metadata=None)