diff --git a/concrete/quantization/post_training.py b/concrete/quantization/post_training.py
index 0035e258a..10a61ce5a 100644
--- a/concrete/quantization/post_training.py
+++ b/concrete/quantization/post_training.py
@@ -91,7 +91,12 @@ class PostTrainingAffineQuantization:
                 # Check if layer is last layer from the model
                 if name == list(self.numpy_model.torch_model.named_children())[-1][0]:
                     # If last layer, we can use 7 bits (maximum allowed) of precision.
-                    q_layer = QuantizedLinear(7, q_weights, q_bias)
+                    # However, 6 bits is currently used to allow 100% FHE precision
+                    # compared to its quantized counterpart.
+                    # Since this is the last layer and mostly used for classification,
+                    # this does not have much impact.
+                    # Put back 7 bits when 100% at 7b is achieved.
+                    q_layer = QuantizedLinear(6, q_weights, q_bias)
                 else:
                     q_layer = QuantizedLinear(self.n_bits, q_weights, q_bias)
                 # Calibrate and get new calibration_data for next layer/activation
diff --git a/tests/torch/test_compile_torch.py b/tests/torch/test_compile_torch.py
index f6f6b036d..b3090cc9b 100644
--- a/tests/torch/test_compile_torch.py
+++ b/tests/torch/test_compile_torch.py
@@ -8,7 +8,7 @@ from concrete.torch.compile import compile_torch_model
 
 # INPUT_OUTPUT_FEATURE is the number of input and output of each of the network layers.
 # (as well as the input of the network itself)
-INPUT_OUTPUT_FEATURE = [1, 2, 3]
+INPUT_OUTPUT_FEATURE = [1, 2]
 
 
 class FC(nn.Module):