diff --git a/core/general-client/src/client.cpp b/core/general-client/src/client.cpp
index 4d3b99f2d8c00fd8dace85b219ce60b2b7444ff5..cc55dd30a5649afac98810fb83f98a837932a523 100644
--- a/core/general-client/src/client.cpp
+++ b/core/general-client/src/client.cpp
@@ -23,8 +23,7 @@ using configure::GeneralModelConfig;
 using baidu::paddle_serving::predictor::general_model::Request;
 using baidu::paddle_serving::predictor::general_model::Response;
 using baidu::paddle_serving::predictor::general_model::Tensor;
-// paddle inference 2.1 support: FLOAT32, INT64, INT32, UINT8, INT8
-// will support: FLOAT16
+// support: FLOAT32, INT64, INT32, UINT8, INT8, FLOAT16
 enum ProtoDataType {
   P_INT64 = 0,
   P_FLOAT32,
@@ -431,7 +430,8 @@ int PredictorOutputs::ParseProto(const Response& res,
             output.tensor(idx).int_data().begin(),
             output.tensor(idx).int_data().begin() + size);
       } else if (fetch_name_to_type[name] == P_UINT8
-                || fetch_name_to_type[name] == P_INT8) {
+                || fetch_name_to_type[name] == P_INT8
+                || fetch_name_to_type[name] == P_FP16) {
         VLOG(2) << "fetch var [" << name << "]type="
                 << fetch_name_to_type[name];
         string_data_map[name] = output.tensor(idx).tensor_content();
diff --git a/core/general-client/src/general_model.cpp b/core/general-client/src/general_model.cpp
index fb71c0c9fc6e3680b8b51bad9ca891e41ef3a849..403119594c759a35d5dfd6251174627f367d9c65 100644
--- a/core/general-client/src/general_model.cpp
+++ b/core/general-client/src/general_model.cpp
@@ -25,8 +25,7 @@ using baidu::paddle_serving::Timer;
 using baidu::paddle_serving::predictor::general_model::Request;
 using baidu::paddle_serving::predictor::general_model::Response;
 using baidu::paddle_serving::predictor::general_model::Tensor;
-// paddle inference support: FLOAT32, INT64, INT32, UINT8, INT8
-// will support: FLOAT16
+// support: FLOAT32, INT64, INT32, UINT8, INT8, FLOAT16
 enum ProtoDataType {
   P_INT64 = 0,
   P_FLOAT32,
diff --git a/core/general-server/op/general_reader_op.cpp b/core/general-server/op/general_reader_op.cpp
index 23e4cc66e96be2c71f12f4ccc1df5f3b8a7f137c..2ad3e4cab6b77b305494c3833f0e3781ed0fd0b7 100644
--- a/core/general-server/op/general_reader_op.cpp
+++ b/core/general-server/op/general_reader_op.cpp
@@ -31,8 +31,7 @@ using baidu::paddle_serving::predictor::MempoolWrapper;
 using baidu::paddle_serving::predictor::general_model::Tensor;
 using baidu::paddle_serving::predictor::general_model::Request;
 using baidu::paddle_serving::predictor::PaddleGeneralModelConfig;
-// paddle inference 2.1 support: FLOAT32, INT64, INT32, UINT8, INT8
-// will support: FLOAT16
+// support: FLOAT32, INT64, INT32, UINT8, INT8, FLOAT16
 enum ProtoDataType {
   P_INT64 = 0,
   P_FLOAT32,
diff --git a/python/paddle_serving_client/client.py b/python/paddle_serving_client/client.py
index 9a8bb3db0180b8ce4617aeee6c7462da490884d8..648678f3afd9ffdc0af4c505779fc5eca0c42a37 100755
--- a/python/paddle_serving_client/client.py
+++ b/python/paddle_serving_client/client.py
@@ -551,6 +551,22 @@ class Client(object):
                         tmp_lod = result_batch_handle.get_lod(mi, name)
                         if np.size(tmp_lod) > 0:
                             result_map["{}.lod".format(name)] = tmp_lod
+                elif self.fetch_names_to_type_[name] == float16_type:
+                    # result_map[name] will be py::array(numpy array)
+                    tmp_str = result_batch_handle.get_string_by_name(
+                        mi, name)
+                    result_map[name] = np.fromstring(tmp_str, dtype = np.float16)
+                    if result_map[name].size == 0:
+                        raise ValueError(
+                            "Failed to fetch, maybe the type of [{}]"
+                            " is wrong, please check the model file".format(
+                                name))
+                    shape = result_batch_handle.get_shape(mi, name)
+                    result_map[name].shape = shape
+                    if name in self.lod_tensor_set:
+                        tmp_lod = result_batch_handle.get_lod(mi, name)
+                        if np.size(tmp_lod) > 0:
+                            result_map["{}.lod".format(name)] = tmp_lod
             multi_result_map.append(result_map)
         ret = None
         if len(model_engine_names) == 1: