diff --git a/paddle/fluid/operators/mean_op.cc b/paddle/fluid/operators/mean_op.cc
index 764529a15b6a2b2c98f9ac727d971b8b0b8d1855..83fe1aa6dd1482bfa2a3ba99f488e4ec66d2b49f 100644
--- a/paddle/fluid/operators/mean_op.cc
+++ b/paddle/fluid/operators/mean_op.cc
@@ -96,7 +96,11 @@ REGISTER_OPERATOR(mean_grad, ops::MeanGradOp,
                   ops::MeanGradNoNeedBufferVarsInferer);
 REGISTER_OP_CPU_KERNEL(
     mean, ops::MeanKernel<paddle::platform::CPUDeviceContext, float>,
-    ops::MeanKernel<paddle::platform::CPUDeviceContext, double>);
+    ops::MeanKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::MeanKernel<paddle::platform::CPUDeviceContext,
+                    paddle::platform::bfloat16>);
 REGISTER_OP_CPU_KERNEL(
     mean_grad, ops::MeanGradKernel<paddle::platform::CPUDeviceContext, float>,
-    ops::MeanGradKernel<paddle::platform::CPUDeviceContext, double>);
+    ops::MeanGradKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::MeanGradKernel<paddle::platform::CPUDeviceContext,
+                        paddle::platform::bfloat16>);
diff --git a/paddle/pten/kernels/cpu/math.cc b/paddle/pten/kernels/cpu/math.cc
index 25c4671baad7c612c6d6133c4bb1e4ebbbe8d5e9..d63292ba1f5de7bee6f13550f54a447e3279b17b 100644
--- a/paddle/pten/kernels/cpu/math.cc
+++ b/paddle/pten/kernels/cpu/math.cc
@@ -97,7 +97,8 @@ using complex128 = ::paddle::platform::complex<double>;
 // using bfloat16 = ::paddle::platform::bfloat16;
 
 PT_REGISTER_KERNEL("sign", CPU, ANY, pten::Sign, float, double) {}
-PT_REGISTER_KERNEL("mean", CPU, ANY, pten::Mean, float, double) {}
+PT_REGISTER_KERNEL(
+    "mean", CPU, ANY, pten::Mean, float, double, paddle::platform::bfloat16) {}
 PT_REGISTER_KERNEL("scale",
                    CPU,
                    ANY,
diff --git a/python/paddle/fluid/tests/book/test_fit_a_line.py b/python/paddle/fluid/tests/book/test_fit_a_line.py
index 8db8b793597bd57e9343af9e37f3cf5150d96225..a8a5c8bf31598c50edd25d8147ce115eb44df985 100644
--- a/python/paddle/fluid/tests/book/test_fit_a_line.py
+++ b/python/paddle/fluid/tests/book/test_fit_a_line.py
@@ -24,10 +24,19 @@ import unittest
 import math
 import sys
 import os
+import struct
 
 paddle.enable_static()
 
 
+def convert_uint16_to_float(in_list):
+    in_list = numpy.asarray(in_list)
+    out = numpy.vectorize(
+        lambda x: struct.unpack('<f', struct.pack('<I', x << 16))[0],
+        otypes=[numpy.float32])(in_list.flat)
+    return numpy.reshape(out, in_list.shape)
+
+
 def train(use_cuda, save_dirname, is_local, use_bf16, pure_bf16):
     x = fluid.layers.data(name='x', shape=[13], dtype='float32')
     y = fluid.layers.data(name='y', shape=[1], dtype='float32')
@@ -84,6 +93,8 @@ def train(use_cuda, save_dirname, is_local, use_bf16, pure_bf16):
                 avg_loss_value, = exe.run(main_program,
                                           feed=feeder.feed(data),
                                           fetch_list=[avg_cost])
+                if avg_loss_value.dtype == numpy.uint16:
+                    avg_loss_value = convert_uint16_to_float(avg_loss_value)
                 if avg_loss_value[0] < 10.0:
                     if save_dirname is not None:
                         paddle.static.save_inference_model(
@@ -154,6 +165,8 @@ def infer(use_cuda, save_dirname=None, use_bf16=False):
         results = exe.run(inference_program,
                           feed={feed_target_names[0]: numpy.array(test_feat)},
                           fetch_list=fetch_targets)
+        if results[0].dtype == numpy.uint16:
+            results[0] = convert_uint16_to_float(results[0])
         print("infer shape: ", results[0].shape)
         print("infer results: ", results[0])
         print("ground truth: ", test_label)
diff --git a/python/paddle/fluid/tests/unittests/test_mean_op.py b/python/paddle/fluid/tests/unittests/test_mean_op.py
index d5cc81456b84b179fa4b29088147e0a75313c1dc..064de404aff27a5cc5fc210ebb82a06c516a9f81 100644
--- a/python/paddle/fluid/tests/unittests/test_mean_op.py
+++ b/python/paddle/fluid/tests/unittests/test_mean_op.py
@@ -16,7 +16,7 @@ from __future__ import print_function
 
 import unittest
 import numpy as np
-from op_test import OpTest
+from op_test import OpTest, OpTestTool
 import paddle
 import paddle.fluid.core as core
 import paddle.fluid as fluid
@@ -76,6 +76,20 @@ class TestFP16MeanOp(TestMeanOp):
                 place, ['X'], 'Out', max_relative_error=0.8)
 
 
+@OpTestTool.skip_if_not_cpu_bf16()
+class TestBF16MeanOp(TestMeanOp):
+    def init_dtype_type(self):
+        self.dtype = np.uint16
+
+    def test_check_output(self):
+        paddle.enable_static()
+        self.check_output_with_place(core.CPUPlace())
+
+    def test_checkout_grad(self):
+        place = core.CPUPlace()
+        self.check_grad_with_place(place, ['X'], 'Out')
+
+
 def ref_reduce_mean(x, axis=None, keepdim=False, reduce_all=False):
     if isinstance(axis, list):
         axis = tuple(axis)