diff --git a/paddle/fluid/lite/CMakeLists.txt b/paddle/fluid/lite/CMakeLists.txt
index 542a70a5fe1aa6abc303d4c24eff67c0035184b8..7b6dd0703d410ad228a11e60dda7ceea9f5a7983 100644
--- a/paddle/fluid/lite/CMakeLists.txt
+++ b/paddle/fluid/lite/CMakeLists.txt
@@ -11,7 +11,7 @@ message(STATUS "LITE_WITH_PROFILE:\t${LITE_WITH_PROFILE}")
 
 set(LITE_MODEL_DIR "${THIRD_PARTY_PATH}/install")
 
-set(LITE_ON_MOBILE LITE_WITH_LIGHT_WEIGHT_FRAMEWORK)
+set(LITE_ON_MOBILE ${LITE_WITH_LIGHT_WEIGHT_FRAMEWORK})
 
 set(LITE_URL "http://paddle-inference-dist.bj.bcebos.com" CACHE STRING "inference download url")
 
@@ -189,4 +189,7 @@ add_subdirectory(gen_code)
 
 if (WITH_TESTING)
     lite_download_and_uncompress(${LITE_MODEL_DIR} ${LITE_URL} "lite_naive_model.tar.gz")
+    if(LITE_WITH_LIGHT_WEIGHT_FRAMEWORK)
+        lite_download_and_uncompress(${LITE_MODEL_DIR} ${LITE_URL} "mobilenet_v2_relu.tar.gz")
+    endif()
 endif()
diff --git a/paddle/fluid/lite/api/CMakeLists.txt b/paddle/fluid/lite/api/CMakeLists.txt
index 5c70a42d1738e6305a8b3f630820715300c06518..52961d0cc49187fa79e55942a1abaceed9dc2d19 100644
--- a/paddle/fluid/lite/api/CMakeLists.txt
+++ b/paddle/fluid/lite/api/CMakeLists.txt
@@ -33,16 +33,26 @@ include(ExternalProject)
 set(LITE_DEMO_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo" CACHE STRING
         "A path setting inference demo download directories.")
 
-if((NOT LITE_WITH_LIGHT_WEIGHT_FRAMEWORK) AND WITH_TESTING)
+if(WITH_TESTING)
+    set(eval_model_dir "")
+    set(test_cxx_api_deps cxx_api_lite mir_passes ${ops_lite} ${host_kernels} ${x86_kernels})
+
+    if(LITE_WITH_LIGHT_WEIGHT_FRAMEWORK)
+        set(eval_model_dir ${LITE_MODEL_DIR}/mobilenet_v2_relu)
+        set(test_cxx_api_deps ${test_cxx_api_deps} ${arm_kernels})
+    endif()
     lite_cc_test(test_cxx_api_lite SRCS cxx_api_test.cc
-       DEPS cxx_api_lite mir_passes
-       ${ops_lite} ${host_kernels} ${x86_kernels}
+       DEPS ${test_cxx_api_deps}
        ARGS --model_dir=${LITE_MODEL_DIR}/lite_naive_model
-            --optimized_model=${LITE_MODEL_DIR}/lite_naive_model_opt SERIAL)
+            --optimized_model=${LITE_MODEL_DIR}/lite_naive_model_opt 
+            --eval_model_dir=eval_model_dir SERIAL)
+
     add_dependencies(test_cxx_api_lite extern_lite_download_lite_naive_model_tar_gz)
+    if(LITE_WITH_LIGHT_WEIGHT_FRAMEWORK)
+        add_dependencies(test_cxx_api_lite extern_lite_download_mobilenet_v2_relu_tar_gz)
+    endif()
 endif()
 
-
 # These tests needs CLI arguments, and is not supported in ARM CI.
 # TODO(Superjomn) support latter.
 if(NOT LITE_ON_MOBILE)
diff --git a/paddle/fluid/lite/api/cxx_api_test.cc b/paddle/fluid/lite/api/cxx_api_test.cc
index 11ae1f9ca804402cd3d7a9bb3e0c44217f61cff5..1b337c06a981447fd8b8f87905ce5d3d10c56d8c 100644
--- a/paddle/fluid/lite/api/cxx_api_test.cc
+++ b/paddle/fluid/lite/api/cxx_api_test.cc
@@ -19,7 +19,6 @@
 #include "paddle/fluid/lite/api/lite_api_test_helper.h"
 #include "paddle/fluid/lite/core/compatible_tensor.h"
 #include "paddle/fluid/lite/core/mir/use_passes.h"
-#include "paddle/fluid/lite/core/mir/use_passes.h"
 #include "paddle/fluid/lite/core/op_registry.h"
 #include "paddle/fluid/lite/kernels/use_kernels.h"
 #include "paddle/fluid/lite/operators/use_ops.h"
@@ -28,9 +27,13 @@
 DEFINE_string(startup_program_path, "", "");
 DEFINE_string(main_program_path, "", "");
 
+// for eval
+DEFINE_string(eval_model_dir, "", "");
+
 namespace paddle {
 namespace lite {
 
+#ifndef LITE_WITH_LIGHT_WEIGHT_FRAMEWORK
 TEST(CXXApi, test) {
   const lite::Tensor* out = RunHvyModel();
   LOG(INFO) << out << " memory size " << out->data_size();
@@ -41,7 +44,6 @@ TEST(CXXApi, test) {
   // LOG(INFO) << "out " << *out;
 }
 
-#ifndef LITE_WITH_LIGHT_WEIGHT_FRAMEWORK
 TEST(CXXApi, save_model) {
   lite::ExecutorLite predictor;
   std::vector<Place> valid_places({Place{TARGET(kHost), PRECISION(kFloat)},
@@ -52,9 +54,7 @@ TEST(CXXApi, save_model) {
   LOG(INFO) << "Save optimized model to " << FLAGS_optimized_model;
   predictor.SaveModel(FLAGS_optimized_model);
 }
-#endif  // LITE_WITH_LIGHT_WEIGHT_FRAMEWORK
 
-#ifndef LITE_WITH_LIGHT_WEIGHT_FRAMEWORK
 /*TEST(CXXTrainer, train) {
   Place prefer_place({TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW)});
   std::vector<Place> valid_places({prefer_place});
@@ -88,5 +88,37 @@ TEST(CXXApi, save_model) {
 }*/
 #endif  // LITE_WITH_LIGHT_WEIGHT_FRAMEWORK
 
+#ifdef LITE_WITH_ARM
+TEST(CXXApi, eval) {
+  DeviceInfo::Init();
+  lite::ExecutorLite predictor;
+  std::vector<Place> valid_places({Place{TARGET(kHost), PRECISION(kFloat)},
+                                   Place{TARGET(kARM), PRECISION(kFloat)}});
+
+  predictor.Build(FLAGS_eval_model_dir, Place{TARGET(kARM), PRECISION(kFloat)},
+                  valid_places);
+
+  auto* input_tensor = predictor.GetInput(0);
+  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 224, 224})));
+  auto* data = input_tensor->mutable_data<float>();
+  for (int i = 0; i < input_tensor->dims().production(); i++) {
+    data[i] = 1;
+  }
+
+  predictor.Run();
+
+  auto* out = predictor.GetOutput(0);
+  std::vector<float> results({0.00097802, 0.00099822, 0.00103093, 0.00100121,
+                              0.00098268, 0.00104065, 0.00099962, 0.00095181,
+                              0.00099694, 0.00099406});
+  for (int i = 0; i < results.size(); ++i) {
+    EXPECT_NEAR(out->data<float>()[i], results[i], 1e-5);
+  }
+  ASSERT_EQ(out->dims().size(), 2);
+  ASSERT_EQ(out->dims()[0], 1);
+  ASSERT_EQ(out->dims()[1], 1000);
+}
+#endif
+
 }  // namespace lite
 }  // namespace paddle
diff --git a/paddle/fluid/lite/kernels/use_kernels.h b/paddle/fluid/lite/kernels/use_kernels.h
index f7d17374010ff6564bfe3ffd085f130bd61d2a16..2c06092e3856467c031abaf36c63bd61aef65bae 100644
--- a/paddle/fluid/lite/kernels/use_kernels.h
+++ b/paddle/fluid/lite/kernels/use_kernels.h
@@ -36,6 +36,19 @@ USE_LITE_KERNEL(depthwise_conv2d, kX86, kFloat, kNCHW, def);
 USE_LITE_KERNEL(pool2d, kX86, kFloat, kNCHW, def);
 #endif
 
+#ifdef LITE_WITH_ARM
+USE_LITE_KERNEL(fc, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(mul, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(scale, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(conv2d, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(batch_norm, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(relu, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(depthwise_conv2d, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(pool2d, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(elementwise_add, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(softmax, kARM, kFloat, kNCHW, def);
+#endif
+
 #ifdef LITE_WITH_CUDA
 USE_LITE_KERNEL(mul, kCUDA, kFloat, kNCHW, def);
 USE_LITE_KERNEL(io_copy, kCUDA, kAny, kAny, host_to_device);
diff --git a/paddle/fluid/lite/operators/use_ops.h b/paddle/fluid/lite/operators/use_ops.h
index b747a4659b183faf73c0242eff005b1af2ec6180..8f7599042b5538a9bff248a84c5f3f3980c9500b 100644
--- a/paddle/fluid/lite/operators/use_ops.h
+++ b/paddle/fluid/lite/operators/use_ops.h
@@ -33,3 +33,4 @@ USE_LITE_OP(concat)
 USE_LITE_OP(conv2d)
 USE_LITE_OP(depthwise_conv2d)
 USE_LITE_OP(pool2d)
+USE_LITE_OP(batch_norm)
diff --git a/paddle/fluid/lite/tools/build.sh b/paddle/fluid/lite/tools/build.sh
index 44c546e1de35447f258d051b5ddb73e2ed1b86cf..29fa9d9ad0b6c0b46e41ad12cee615bee4928bcc 100755
--- a/paddle/fluid/lite/tools/build.sh
+++ b/paddle/fluid/lite/tools/build.sh
@@ -114,14 +114,47 @@ function test_arm_android {
 
     echo "test name: ${test_name}"
     adb_work_dir="/data/local/tmp"
-    skip_list="test_model_parser_lite" # add more with space
-    [[ $skip_list =~ (^|[[:space:]])$test_name($|[[:space:]]) ]] && continue || echo 'skip $test_name'
+
+    skip_list=("test_model_parser_lite" "test_cxx_api_lite")
+    for skip_name in ${skip_list[@]} ; do
+        [[ $skip_name =~ (^|[[:space:]])$test_name($|[[:space:]]) ]] && echo "skip $test_name" && return
+    done
+
     testpath=$(find ./paddle/fluid -name ${test_name})
     adb -s emulator-${port} push ${testpath} ${adb_work_dir}
     adb -s emulator-${port} shell chmod +x "${adb_work_dir}/${test_name}"
     adb -s emulator-${port} shell "./${adb_work_dir}/${test_name}"
 }
 
+function test_arm_model {
+    local test_name=$1
+    local port=$2
+    local model_dir=$3
+
+    if [[ "${test_name}x" == "x" ]]; then
+        echo "test_name can not be empty"
+        exit 1
+    fi
+    if [[ "${port}x" == "x" ]]; then
+        echo "Port can not be empty"
+        exit 1
+    fi
+    if [[ "${model_dir}x" == "x" ]]; then
+        echo "Model dir can not be empty"
+        exit 1
+    fi
+
+    echo "test name: ${test_name}"
+    adb_work_dir="/data/local/tmp"
+
+    testpath=$(find ./paddle/fluid -name ${test_name})
+    adb -s emulator-${port} push ${model_dir} ${adb_work_dir}
+    adb -s emulator-${port} push ${testpath} ${adb_work_dir}
+    adb -s emulator-${port} shell chmod +x "${adb_work_dir}/${test_name}"
+    local adb_model_path="./${adb_work_dir}/`basename ${model_dir}`"
+    adb -s emulator-${port} shell "./${adb_work_dir}/${test_name} --eval_model_dir=$adb_model_path"
+}
+
 # Build the code and run lite arm tests. This is executed in the CI system.
 function build_test_arm {
     # 1. Build goes first
@@ -196,6 +229,8 @@ function build_test_arm {
             for _test in $(cat $TESTS_FILE); do
                 test_arm_android $_test $port
             done
+            # TODO(sangoly): refine this
+            test_arm_model "test_cxx_api_lite" $port "./third_party/install/mobilenet_v2_relu"
         done
     done