diff --git a/flax/serialization.py b/flax/serialization.py
index bd6d0853cc..e75e362c89 100644
--- a/flax/serialization.py
+++ b/flax/serialization.py
@@ -263,6 +263,9 @@ def _dtype_from_name(name: str):
   """Handle JAX bfloat16 dtype correctly."""
   if name == b'bfloat16':
     return jax.numpy.bfloat16
+  elif name[0:4] == b'void':
+    # decode voidNNNN, where NNNN is the length in bits
+    return np.void(int(name[4:]) // 8)
   else:
     return np.dtype(name)
 
@@ -343,6 +346,10 @@ def _np_convert_in_place(d):
 
 def _chunk(arr) -> dict[str, Any]:
   """Convert array to a canonical dictionary of chunked arrays."""
+  if isinstance(arr.dtype, np.dtypes.VoidDType):
+    # The chunking strategy below doesn't work for a large scalar np.void.
+    # Implement an alternative strategy later if needed.
+    raise NotImplementedError('Chunking not implemented for np.void dtype')
   chunksize = max(1, int(MAX_CHUNK_SIZE / arr.dtype.itemsize))
   data = {'__msgpack_chunked_array__': True, 'shape': _tuple_to_dict(arr.shape)}
   flatarr = arr.reshape(-1)
diff --git a/tests/serialization_test.py b/tests/serialization_test.py
index 05c319b6e8..ed05d0c486 100644
--- a/tests/serialization_test.py
+++ b/tests/serialization_test.py
@@ -264,6 +264,7 @@ def __call__(self):
       'float',
       'complex',
       'bool',
+      'void',
   ])
   def test_numpy_serialization(self, dtype):
     np.random.seed(0)