diff --git a/CHANGELOG.md b/CHANGELOG.md
index cb283378..122eea12 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,12 @@
 # Changelog
 All notable changes to this project will be documented in this file.
 
+### [2.1.12]
+
+#### Fixed
+
+- Fix wrong dtype used when evaluating finetuning or anomaly models trained in fp16 precision
+
 ### [2.1.11]
 
 #### Fixed
diff --git a/pyproject.toml b/pyproject.toml
index 32769a62..0fd7fe3b 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "quadra"
-version = "2.1.11"
+version = "2.1.12"
 description = "Deep Learning experiment orchestration library"
 authors = [
 	"Federico Belotti <federico.belotti@orobix.com>",
diff --git a/quadra/__init__.py b/quadra/__init__.py
index f39f08a9..67acf36f 100644
--- a/quadra/__init__.py
+++ b/quadra/__init__.py
@@ -1,4 +1,4 @@
-__version__ = "2.1.11"
+__version__ = "2.1.12"
 
 
 def get_version():
diff --git a/quadra/tasks/anomaly.py b/quadra/tasks/anomaly.py
index ec617cd9..97f3431a 100644
--- a/quadra/tasks/anomaly.py
+++ b/quadra/tasks/anomaly.py
@@ -386,10 +386,11 @@ def test(self) -> None:
                 batch_labels = batch_item["label"]
                 image_labels.extend(batch_labels.tolist())
                 image_paths.extend(batch_item["image_path"])
+                batch_images = batch_images.to(device=self.device, dtype=self.deployment_model.model_dtype)
                 if self.model_data.get("anomaly_method") == "efficientad":
-                    model_output = self.deployment_model(batch_images.to(self.device), None)
+                    model_output = self.deployment_model(batch_images, None)
                 else:
-                    model_output = self.deployment_model(batch_images.to(self.device))
+                    model_output = self.deployment_model(batch_images)
                 anomaly_map, anomaly_score = model_output[0], model_output[1]
                 anomaly_map = anomaly_map.cpu()
                 anomaly_score = anomaly_score.cpu()
diff --git a/quadra/tasks/classification.py b/quadra/tasks/classification.py
index eb05fb52..e885c2c6 100644
--- a/quadra/tasks/classification.py
+++ b/quadra/tasks/classification.py
@@ -1169,7 +1169,7 @@ def test(self) -> None:
         with torch.set_grad_enabled(self.gradcam):
             for batch_item in tqdm(test_dataloader):
                 im, target = batch_item
-                im = im.to(self.device).detach()
+                im = im.to(device=self.device, dtype=self.deployment_model.model_dtype).detach()
 
                 if self.gradcam:
                     # When gradcam is used we need to remove gradients
diff --git a/quadra/tasks/segmentation.py b/quadra/tasks/segmentation.py
index 327ba620..04e68352 100644
--- a/quadra/tasks/segmentation.py
+++ b/quadra/tasks/segmentation.py
@@ -346,10 +346,7 @@ def test(self) -> None:
             image_list, mask_list, mask_pred_list, label_list = [], [], [], []
             for batch in dataloader:
                 images, masks, labels = batch
-                images = images.to(self.device)
-                # TODO: This can be problematic for the future considering bfloat16 or float16-true.
-                if "16" in str(self.deployment_model.model_dtype):
-                    images = images.half()
+                images = images.to(device=self.device, dtype=self.deployment_model.model_dtype)
                 if len(masks.shape) == 3:  # BxHxW -> Bx1xHxW
                     masks = masks.unsqueeze(1)
                 with torch.no_grad():