diff --git a/tueri/output_scanners/factual_consistency.py b/tueri/output_scanners/factual_consistency.py
index 1cb7bb63..e3f6e726 100644
--- a/tueri/output_scanners/factual_consistency.py
+++ b/tueri/output_scanners/factual_consistency.py
@@ -50,7 +50,11 @@ def __init__(
             use_onnx=use_onnx,
         )
         if not use_onnx:
-            self._model = self._model.to(device())
+            # Handle meta device properly - use to_empty() when moving from meta device
+            if hasattr(self._model, 'device') and str(self._model.device) == 'meta':
+                self._model = self._model.to_empty(device=device())
+            else:
+                self._model = self._model.to(device())
             self._model.eval()
 
     def scan(self, prompt: str, output: str) -> tuple[str, bool, float]:
diff --git a/tueri/output_scanners/relevance.py b/tueri/output_scanners/relevance.py
index c5715a37..402da9c9 100644
--- a/tueri/output_scanners/relevance.py
+++ b/tueri/output_scanners/relevance.py
@@ -110,7 +110,12 @@ def __init__(
                 subfolder=model.subfolder,
                 revision=model.revision,
                 **model.kwargs,
-            ).to(device())
+            )
+            # Handle meta device properly - use to_empty() when moving from meta device
+            if hasattr(self._model, 'device') and str(self._model.device) == 'meta':
+                self._model = self._model.to_empty(device=device())
+            else:
+                self._model = self._model.to(device())
             LOGGER.debug("Initialized model", model=model, device=device())
             self._model.eval()
 
diff --git a/tueri/transformers_helpers.py b/tueri/transformers_helpers.py
index 8fa34edb..1cf58632 100644
--- a/tueri/transformers_helpers.py
+++ b/tueri/transformers_helpers.py
@@ -93,7 +93,11 @@ def get_tokenizer_and_model_for_classification(
             low_cpu_mem_usage=False,
             **model.kwargs,
         )
-        tf_model = tf_model.to(device())
+        # Handle meta device properly - use to_empty() when moving from meta device
+        if hasattr(tf_model, 'device') and str(tf_model.device) == 'meta':
+            tf_model = tf_model.to_empty(device=device())
+        else:
+            tf_model = tf_model.to(device())
         LOGGER.debug("Initialized classification model", model=model, device=device())
 
         return tf_tokenizer, tf_model
@@ -131,7 +135,11 @@ def get_tokenizer_and_model_for_ner(
             low_cpu_mem_usage=False,
             **model.kwargs,
         )
-        tf_model = tf_model.to(device())
+        # Handle meta device properly - use to_empty() when moving from meta device
+        if hasattr(tf_model, 'device') and str(tf_model.device) == 'meta':
+            tf_model = tf_model.to_empty(device=device())
+        else:
+            tf_model = tf_model.to(device())
         LOGGER.debug("Initialized NER model", model=model, device=device())
 
         return tf_tokenizer, tf_model