diff --git a/deeptab/architectures/mambatab.py b/deeptab/architectures/mambatab.py
index a70705f..5aa75df 100644
--- a/deeptab/architectures/mambatab.py
+++ b/deeptab/architectures/mambatab.py
@@ -110,6 +110,7 @@ def forward(self, *data):
 
         x = self.norm_f(x)
         x = self.embedding_activation(x)
+        x = self.mamba(x)
         if self.axis == 1:
             x = x.squeeze(1)
         else:
diff --git a/tests/test_mambatab_backbone.py b/tests/test_mambatab_backbone.py
new file mode 100644
index 0000000..2eeeabb
--- /dev/null
+++ b/tests/test_mambatab_backbone.py
@@ -0,0 +1,36 @@
+"""Regression test: MambaTab must actually run its Mamba block.
+
+forward() went initial_layer -> norm -> activation -> head and never called
+self.mamba, so the model trained as a linear layer with an MLP head while all
+Mamba parameters sat in the optimizer with no gradients.
+"""
+
+from typing import Any, cast
+
+import numpy as np
+import pandas as pd
+
+from deeptab.models import MambaTabRegressor
+
+
+def _backward_one_batch(model) -> Any:
+    task = cast(Any, model._task_model)
+    task.zero_grad()
+    (num, cat, emb), labels = next(iter(cast(Any, model._data_module).train_dataloader()))
+    preds = task(num, cat, emb)
+    task.compute_loss(preds, labels).backward()
+    return task
+
+
+def test_mambatab_trains_its_mamba_block():
+    rng = np.random.RandomState(0)
+    X = pd.DataFrame({"num1": rng.randn(60), "num2": rng.rand(60)})
+    y = rng.randn(60)
+
+    model = MambaTabRegressor()
+    model.fit(X, y, max_epochs=1, batch_size=16, accelerator="cpu")
+
+    task = _backward_one_batch(model)
+    mamba_params = [(name, p) for name, p in task.named_parameters() if ".mamba." in name]
+    assert mamba_params, "expected mamba parameters on the estimator"
+    assert all(p.grad is not None for _, p in mamba_params)