diff --git a/torchmdnet/models/tensornet.py b/torchmdnet/models/tensornet.py
index 89b1cfe4c..956f683bd 100644
--- a/torchmdnet/models/tensornet.py
+++ b/torchmdnet/models/tensornet.py
@@ -55,34 +55,31 @@ def tensor_norm(tensor):
 
 
 class TensorNet(nn.Module):
-    r"""TensorNet's architecture.
+    r"""TensorNet's architecture, from TensorNet: Cartesian Tensor Representations
+        for Efficient Learning of Molecular Potentials; G. Simeon and G. de Fabritiis.
 
     Args:
         hidden_channels (int, optional): Hidden embedding size.
             (default: :obj:`128`)
         num_layers (int, optional): The number of interaction layers.
-            (default: :obj:`6`)
+            (default: :obj:`2`)
         num_rbf (int, optional): The number of radial basis functions :math:`\mu`.
-            (default: :obj:`50`)
+            (default: :obj:`32`)
         rbf_type (string, optional): The type of radial basis function to use.
             (default: :obj:`"expnorm"`)
         trainable_rbf (bool, optional): Whether to train RBF parameters with
-            backpropagation. (default: :obj:`True`)
+            backpropagation. (default: :obj:`False`)
         activation (string, optional): The type of activation function to use.
             (default: :obj:`"silu"`)
         cutoff_lower (float, optional): Lower cutoff distance for interatomic interactions.
             (default: :obj:`0.0`)
         cutoff_upper (float, optional): Upper cutoff distance for interatomic interactions.
-            (default: :obj:`5.0`)
+            (default: :obj:`4.5`)
         max_z (int, optional): Maximum atomic number. Used for initializing embeddings.
-            (default: :obj:`100`)
+            (default: :obj:`128`)
         max_num_neighbors (int, optional): Maximum number of neighbors to return for a
             given node/atom when constructing the molecular graph during forward passes.
-            This attribute is passed to the torch_cluster radius_graph routine keyword
-            max_num_neighbors, which normally defaults to 32. Users should set this to
-            higher values if they are using higher upper distance cutoffs and expect more
-            than 32 neighbors per node/atom.
-            (default: :obj:`32`)
+            (default: :obj:`64`)
         equivariance_invariance_group (string, optional): Group under whose action on input
             positions internal tensor features will be equivariant and scalar predictions
             will be invariant. O(3) or SO(3).