diff --git a/src/core/serialization.rs b/src/core/serialization.rs
index 62b5a5da..619daba6 100644
--- a/src/core/serialization.rs
+++ b/src/core/serialization.rs
@@ -120,11 +120,10 @@ impl Loader for SafeTensorLoader {
                         {
                             // Convert to fp32
                             let data: Vec<f32> = match tensor_view.dtype() {
-                                Dtype::F32 => tensor_view
-                                    .data()
-                                    .chunks_exact(4)
-                                    .map(|c| f32::from_ne_bytes([c[0], c[1], c[2], c[3]]))
-                                    .collect(),
+                                Dtype::F32 => {
+                                    unsafe { std::mem::transmute::<_, &[f32]>(tensor_view.data()) }
+                                        .to_vec()
+                                }
                                 Dtype::F16 => tensor_view
                                     .data()
                                     .chunks_exact(2)
@@ -199,18 +198,8 @@ impl<'data> View for &'data Tensor {
 
 impl<'a> std::convert::From<safetensors::tensor::TensorView<'a>> for Tensor {
     fn from(value: safetensors::tensor::TensorView<'a>) -> Self {
-        let chunked = value.data().chunks_exact(std::mem::size_of::<f32>());
-
         Tensor {
-            data: Box::new(
-                chunked
-                    .map(|chunk| unsafe {
-                        std::mem::transmute::<[u8; 4], f32>([
-                            chunk[0], chunk[1], chunk[2], chunk[3],
-                        ])
-                    })
-                    .collect::<Vec<f32>>(),
-            ),
+            data: Box::new(unsafe { std::mem::transmute::<_, &'a [f32]>(value.data()) }.to_vec()),
         }
     }
 }