diff --git a/src/fabio/compression/compression.py b/src/fabio/compression/compression.py
index 8fa668988..48133b596 100644
--- a/src/fabio/compression/compression.py
+++ b/src/fabio/compression/compression.py
@@ -37,7 +37,7 @@
 __author__ = "Jérôme Kieffer"
 __contact__ = "jerome.kieffer@esrf.eu"
 __license__ = "MIT"
-__date__ = "06/04/2020"
+__date__ = "20/10/2023"
 __copyright__ = "European Synchrotron Radiation Facility, Grenoble, France"
 
 import sys
@@ -205,6 +205,7 @@ def decByteOffset_numpy(stream, size=None, dtype="int64"):
     :param size: the size of the output array (of longInts)
     :return: 1D-ndarray
 
+    Nota: data are always stored as little endian.
     """
     logger.debug("CBF decompression using Numpy")
     listnpa = []
@@ -222,23 +223,19 @@ def decByteOffset_numpy(stream, size=None, dtype="int64"):
         if stream[idx + 1:idx + 3] == key32:
             if stream[idx + 3:idx + 7] == key64:
                 # 64 bits int
-                res = numpy.frombuffer(stream[idx + 7:idx + 15], dtype=numpy.int64)
+                res = numpy.frombuffer(stream[idx + 7:idx + 15], dtype="<q")
                 listnpa.append(res)
                 shift = 15
             else:
                 # 32 bits int
-                res = numpy.frombuffer(stream[idx + 3:idx + 7], dtype=numpy.int32)
+                res = numpy.frombuffer(stream[idx + 3:idx + 7], dtype="<i")
                 listnpa.append(res)
                 shift = 7
         else:  # int16
-            res = numpy.frombuffer(stream[idx + 1:idx + 3], dtype=numpy.int16)
+            res = numpy.frombuffer(stream[idx + 1:idx + 3], dtype="<h")
             listnpa.append(res)
             shift = 3
         stream = stream[idx + shift:]
-    if not numpy.little_endian:
-        for res in listnpa:
-            if res.dtype != numpy.int8:
-                res.byteswap(True)
     return numpy.ascontiguousarray(numpy.hstack(listnpa), dtype).cumsum()