diff --git a/botorch/models/gpytorch.py b/botorch/models/gpytorch.py
index 414e7ebe64..0c331c1de7 100644
--- a/botorch/models/gpytorch.py
+++ b/botorch/models/gpytorch.py
@@ -579,7 +579,7 @@ def posterior(
         with gpt_posterior_settings():
             # only compute what's necessary
             if output_indices is not None:
-                mvns = [self.forward_i(i, transformed_X[i]) for i in output_indices]
+                mvns = [self.models[i](transformed_X[i]) for i in output_indices]
                 if observation_noise is not False:
                     if torch.is_tensor(observation_noise):
                         lh_kwargs = [
diff --git a/test/models/test_gpytorch.py b/test/models/test_gpytorch.py
index 12660fd130..6ac7bd9189 100644
--- a/test/models/test_gpytorch.py
+++ b/test/models/test_gpytorch.py
@@ -458,9 +458,25 @@ def test_model_list_gpytorch_model(self):
             self.assertIsInstance(posterior, GPyTorchPosterior)
             self.assertEqual(posterior.mean.shape, torch.Size([2, 2]))
             # test output indices
-            posterior = model.posterior(test_X, output_indices=[0])
-            self.assertIsInstance(posterior, GPyTorchPosterior)
-            self.assertEqual(posterior.mean.shape, torch.Size([2, 1]))
+            for output_indices in ([0], [1], [0, 1]):
+                posterior_subset = model.posterior(
+                    test_X, output_indices=output_indices
+                )
+                self.assertIsInstance(posterior_subset, GPyTorchPosterior)
+                self.assertEqual(
+                    posterior_subset.mean.shape, torch.Size([2, len(output_indices)])
+                )
+                self.assertTrue(
+                    torch.equal(
+                        posterior_subset.mean, posterior.mean[..., output_indices]
+                    )
+                )
+                self.assertTrue(
+                    torch.equal(
+                        posterior_subset.variance,
+                        posterior.variance[..., output_indices],
+                    )
+                )
             # test observation noise
             posterior = model.posterior(test_X, observation_noise=True)
             self.assertIsInstance(posterior, GPyTorchPosterior)