diff --git a/lite/api/CMakeLists.txt b/lite/api/CMakeLists.txt
index e9f64d19edfec762d8cecbbaa0702171eec814fb..f7f74ab5822a1305e3e8d24cf36a0a458a6494ff 100644
--- a/lite/api/CMakeLists.txt
+++ b/lite/api/CMakeLists.txt
@@ -296,10 +296,10 @@ if (LITE_ON_TINY_PUBLISH)
 endif()
 
 if (LITE_ON_MODEL_OPTIMIZE_TOOL)
-    message(STATUS "Compiling OPT")
-    lite_cc_binary(OPT SRCS opt.cc cxx_api_impl.cc paddle_api.cc cxx_api.cc
+    message(STATUS "Compiling opt")
+    lite_cc_binary(opt SRCS opt.cc cxx_api_impl.cc paddle_api.cc cxx_api.cc
         DEPS gflags kernel op optimizer mir_passes utils)
-    add_dependencies(OPT op_list_h kernel_list_h all_kernel_faked_cc supported_kernel_op_info_h)
+    add_dependencies(opt op_list_h kernel_list_h all_kernel_faked_cc supported_kernel_op_info_h)
 endif(LITE_ON_MODEL_OPTIMIZE_TOOL)
 
 lite_cc_test(test_paddle_api SRCS paddle_api_test.cc DEPS paddle_api_full paddle_api_light
diff --git a/lite/api/android/jni/native/convert_util_jni.h b/lite/api/android/jni/native/convert_util_jni.h
index 5e5d3723e43eb311f64b85f7507a12497d724109..e4adafdc572fdc937f568508aa9d43eb78470d0d 100644
--- a/lite/api/android/jni/native/convert_util_jni.h
+++ b/lite/api/android/jni/native/convert_util_jni.h
@@ -181,6 +181,7 @@ inline MobileConfig jmobileconfig_to_cpp_mobileconfig(JNIEnv *env,
   MobileConfig config;
 
   // set model dir
+  // NOTE: This is a deprecated API and will be removed in latter release.
   jmethodID model_dir_method = env->GetMethodID(
       mobileconfig_jclazz, "getModelDir", "()Ljava/lang/String;");
   jstring java_model_dir =
@@ -190,6 +191,27 @@ inline MobileConfig jmobileconfig_to_cpp_mobileconfig(JNIEnv *env,
     config.set_model_dir(cpp_model_dir);
   }
 
+  // set model from file
+  jmethodID model_file_method = env->GetMethodID(
+      mobileconfig_jclazz, "getModelFromFile", "()Ljava/lang/String;");
+  jstring java_model_file =
+      (jstring)env->CallObjectMethod(jmobileconfig, model_file_method);
+  if (java_model_file != nullptr) {
+    std::string cpp_model_file = jstring_to_cpp_string(env, java_model_file);
+    config.set_model_from_file(cpp_model_file);
+  }
+
+  // set model from buffer
+  jmethodID model_buffer_method = env->GetMethodID(
+      mobileconfig_jclazz, "getModelFromBuffer", "()Ljava/lang/String;");
+  jstring java_model_buffer =
+      (jstring)env->CallObjectMethod(jmobileconfig, model_buffer_method);
+  if (java_model_buffer != nullptr) {
+    std::string cpp_model_buffer =
+        jstring_to_cpp_string(env, java_model_buffer);
+    config.set_model_from_buffer(cpp_model_buffer);
+  }
+
   // set threads
   jmethodID threads_method =
       env->GetMethodID(mobileconfig_jclazz, "getThreads", "()I");
diff --git a/lite/api/android/jni/src/com/baidu/paddle/lite/MobileConfig.java b/lite/api/android/jni/src/com/baidu/paddle/lite/MobileConfig.java
index 5c71db0c92b344e44ea2927305580de1be293f75..e150f98f22113ef6bcedd5e9882e0bd2a6378c97 100644
--- a/lite/api/android/jni/src/com/baidu/paddle/lite/MobileConfig.java
+++ b/lite/api/android/jni/src/com/baidu/paddle/lite/MobileConfig.java
@@ -64,6 +64,44 @@ public class MobileConfig extends ConfigBase {
         return powerMode.value();
     }
 
+    /**
+     * Set model from file.
+     *
+     * @return
+     */
+    public void setModelFromFile(String modelFile) {
+        this.liteModelFile = modelFile;
+    }
+
+    /**
+     * Returns name of model_file.
+     *  
+     * @return liteModelFile
+     */
+    public String getModelFile() {
+        return liteModelFile;
+    }
+
+    /**
+     * Set model from buffer.
+     *
+     * @return
+     */
+    public void setModelFromBuffer(String modelBuffer) {
+        this.liteModelBuffer = modelBuffer;
+    }
+
+    /**
+     * Returns model buffer
+     *  
+     * @return liteModelBuffer
+     */
+    public String getModelBuffer() {
+        return liteModelBuffer;
+    }
+
     private PowerMode powerMode = PowerMode.LITE_POWER_HIGH;
     private int threads = 1;
+    private String liteModelFile;
+    private String liteModelBuffer;
 }
diff --git a/lite/api/apis_test.cc b/lite/api/apis_test.cc
index ac2c385d53ea0a1785393cd488d115d20c4264f1..bb852297d11a8862460ed6f12e007d727aca9428 100644
--- a/lite/api/apis_test.cc
+++ b/lite/api/apis_test.cc
@@ -62,7 +62,7 @@ TEST(CXXApi_LightApi, optim_model) {
 
 TEST(CXXApi_LightApi, save_and_load_model) {
   lite::Predictor cxx_api;
-  lite::LightPredictor light_api(FLAGS_optimized_model);
+  lite::LightPredictor light_api(FLAGS_optimized_model + ".nb", false);
 
   // CXXAPi
   {
diff --git a/lite/api/benchmark.cc b/lite/api/benchmark.cc
index f0cb6841d5b73ea600b9e2b7e2f055192811b6c3..718dbe44296f2d197efc5b567cf0cc211835d176 100644
--- a/lite/api/benchmark.cc
+++ b/lite/api/benchmark.cc
@@ -116,7 +116,7 @@ void Run(const std::vector<std::vector<int64_t>>& input_shapes,
   lite_api::MobileConfig config;
   config.set_threads(FLAGS_threads);
   config.set_power_mode(static_cast<PowerMode>(FLAGS_power_mode));
-  config.set_model_dir(model_dir);
+  config.set_model_from_file(model_dir + ".nb");
 
   auto predictor = lite_api::CreatePaddlePredictor(config);
 
diff --git a/lite/api/cxx_api.cc b/lite/api/cxx_api.cc
index c1e9fc422450adf96d62c68d622907bd7e15b405..f255d3e31d141987242e313366f65701a8310f76 100644
--- a/lite/api/cxx_api.cc
+++ b/lite/api/cxx_api.cc
@@ -239,7 +239,7 @@ void Predictor::Build(const std::string &model_path,
     case lite_api::LiteModelType::kNaiveBuffer:
       CHECK(!model_path.empty())
           << "NaiveBuffer backend only supported combined param";
-      LoadModelNaive(model_path, scope_.get(), &program_desc_);
+      LoadModelNaiveFromFile(model_path, scope_.get(), &program_desc_);
       break;
     default:
       LOG(FATAL) << "Unknown model type";
diff --git a/lite/api/cxx_api_test.cc b/lite/api/cxx_api_test.cc
index 4d711302cb5880247f4a7b7082185c500b9ad6e9..cdf1e838366f4bcafc1c1c991d8805f115de7345 100644
--- a/lite/api/cxx_api_test.cc
+++ b/lite/api/cxx_api_test.cc
@@ -101,7 +101,7 @@ TEST(CXXApi, save_model) {
 TEST(CXXApi, load_model_naive) {
   lite::Predictor predictor;
   std::vector<Place> valid_places({Place{TARGET(kARM), PRECISION(kFloat)}});
-  predictor.Build(FLAGS_optimized_model + ".naive",
+  predictor.Build(FLAGS_optimized_model + ".naive.nb",
                   "",
                   "",
                   valid_places,
diff --git a/lite/api/light_api.cc b/lite/api/light_api.cc
index 1558e286178b461dc04c4366dc3adca81b2dd9de..29d8f4f29ab822f8c9601bbd63a3626abbbf1818 100644
--- a/lite/api/light_api.cc
+++ b/lite/api/light_api.cc
@@ -18,6 +18,17 @@
 namespace paddle {
 namespace lite {
 
+void LightPredictor::Build(const std::string& lite_model_file,
+                           bool model_from_memory) {
+  if (model_from_memory) {
+    LoadModelNaiveFromMemory(lite_model_file, scope_.get(), &cpp_program_desc_);
+  } else {
+    LoadModelNaiveFromFile(lite_model_file, scope_.get(), &cpp_program_desc_);
+  }
+  BuildRuntimeProgram(cpp_program_desc_);
+  PrepareFeedFetch();
+}
+
 void LightPredictor::Build(const std::string& model_dir,
                            const std::string& model_buffer,
                            const std::string& param_buffer,
diff --git a/lite/api/light_api.h b/lite/api/light_api.h
index d1789a9c98333f6e927ba470717d9227729f2108..aa25ea81c7b62238211f96265a4edc49f2d065a1 100644
--- a/lite/api/light_api.h
+++ b/lite/api/light_api.h
@@ -18,6 +18,7 @@
  */
 #pragma once
 
+#include <algorithm>
 #include <map>
 #include <memory>
 #include <string>
@@ -39,12 +40,22 @@ namespace lite {
  */
 class LITE_API LightPredictor {
  public:
-  LightPredictor(
-      const std::string& model_dir,
-      const std::string& model_buffer = "",
-      const std::string& param_buffer = "",
-      bool model_from_memory = false,
-      lite_api::LiteModelType model_type = lite_api::LiteModelType::kProtobuf) {
+  // constructor function of LightPredictor, `lite_model_file` refers to data in
+  // model file or buffer,`model_from_memory` refers to whther to load model
+  // from memory.
+  LightPredictor(const std::string& lite_model_file,
+                 bool model_from_memory = false) {
+    scope_ = std::make_shared<Scope>();
+    Build(lite_model_file, model_from_memory);
+  }
+
+  // NOTE: This is a deprecated API and will be removed in latter release.
+  LightPredictor(const std::string& model_dir,
+                 const std::string& model_buffer = "",
+                 const std::string& param_buffer = "",
+                 bool model_from_memory = false,
+                 lite_api::LiteModelType model_type =
+                     lite_api::LiteModelType::kNaiveBuffer) {
     scope_ = std::make_shared<Scope>();
     Build(model_dir, model_buffer, param_buffer, model_type, model_from_memory);
   }
@@ -69,6 +80,10 @@ class LITE_API LightPredictor {
   void PrepareFeedFetch();
 
  private:
+  void Build(const std::string& lite_model_file,
+             bool model_from_memory = false);
+
+  // NOTE: This is a deprecated API and will be removed in latter release.
   void Build(
       const std::string& model_dir,
       const std::string& model_buffer,
diff --git a/lite/api/light_api_impl.cc b/lite/api/light_api_impl.cc
index a0ae28df0958403237114a3d4b94031829019339..3965843250abe45c43490bdbb4aaed58915e0908 100644
--- a/lite/api/light_api_impl.cc
+++ b/lite/api/light_api_impl.cc
@@ -23,13 +23,17 @@ namespace lite {
 
 void LightPredictorImpl::Init(const lite_api::MobileConfig& config) {
   // LightPredictor Only support NaiveBuffer backend in publish lib
-  raw_predictor_.reset(
-      new LightPredictor(config.model_dir(),
-                         config.model_buffer(),
-                         config.param_buffer(),
-                         config.model_from_memory(),
-                         lite_api::LiteModelType::kNaiveBuffer));
-
+  if (config.lite_model_file().empty()) {
+    raw_predictor_.reset(
+        new LightPredictor(config.model_dir(),
+                           config.model_buffer(),
+                           config.param_buffer(),
+                           config.model_from_memory(),
+                           lite_api::LiteModelType::kNaiveBuffer));
+  } else {
+    raw_predictor_.reset(new LightPredictor(config.lite_model_file(),
+                                            config.model_from_memory()));
+  }
   mode_ = config.power_mode();
   threads_ = config.threads();
 }
diff --git a/lite/api/model_test.cc b/lite/api/model_test.cc
index cf646d823d97213a4a14573f72a95d1a55169c12..190890da4c109f39cc52ca5209cd952f8937f780 100644
--- a/lite/api/model_test.cc
+++ b/lite/api/model_test.cc
@@ -73,7 +73,7 @@ void Run(const std::vector<std::vector<int64_t>>& input_shapes,
          const int repeat,
          const int warmup_times = 0) {
   lite_api::MobileConfig config;
-  config.set_model_dir(model_dir);
+  config.set_model_from_file(model_dir + ".nb");
   config.set_power_mode(power_mode);
   config.set_threads(thread_num);
 
diff --git a/lite/api/opt.cc b/lite/api/opt.cc
index 249656a5cce1c80823b9fbc1ef2b8d5a7676b4e5..c172169e59ec074b81a07e4fc96cd0363c50a10a 100644
--- a/lite/api/opt.cc
+++ b/lite/api/opt.cc
@@ -17,7 +17,7 @@
 #include <gtest/gtest.h>
 #endif
 // "supported_kernel_op_info.h", "all_kernel_faked.cc" and "kernel_src_map.h"
-// are created automatically during OPT's compiling period
+// are created automatically during opt's compiling period
 #include <iomanip>
 #include "all_kernel_faked.cc"  // NOLINT
 #include "kernel_src_map.h"     // NOLINT
diff --git a/lite/api/paddle_api.cc b/lite/api/paddle_api.cc
index aabb53529221bde53b6b2ee27b2efefee2e6054d..9f071cf7780e27defdd1fcd6be02844618165fb6 100644
--- a/lite/api/paddle_api.cc
+++ b/lite/api/paddle_api.cc
@@ -190,5 +190,27 @@ void ConfigBase::set_threads(int threads) {
 #endif
 }
 
+// set model data in combined format, `set_model_from_file` refers to loading
+// model from file, set_model_from_buffer refers to loading model from memory
+// buffer
+void MobileConfig::set_model_from_file(const std::string &x) {
+  lite_model_file_ = x;
+}
+void MobileConfig::set_model_from_buffer(const std::string &x) {
+  lite_model_file_ = x;
+  model_from_memory_ = true;
+}
+void MobileConfig::set_model_buffer(const char *model_buffer,
+                                    size_t model_buffer_size,
+                                    const char *param_buffer,
+                                    size_t param_buffer_size) {
+  LOG(WARNING) << "warning: `set_model_buffer` will be abandened in "
+                  "release/v3.0.0, new method `set_model_from_buffer(const "
+                  "std::string &x)` is recommended.";
+  model_buffer_ = std::string(model_buffer, model_buffer + model_buffer_size);
+  param_buffer_ = std::string(param_buffer, param_buffer + param_buffer_size);
+  model_from_memory_ = true;
+}
+
 }  // namespace lite_api
 }  // namespace paddle
diff --git a/lite/api/paddle_api.h b/lite/api/paddle_api.h
index 6308699ac91900d161a55ee121e4d9777947fede..307eeb74e8b4cdc3b2d6188eb18490e4dcf89b8f 100644
--- a/lite/api/paddle_api.h
+++ b/lite/api/paddle_api.h
@@ -168,22 +168,40 @@ class LITE_API CxxConfig : public ConfigBase {
 /// MobileConfig is the config for the light weight predictor, it will skip
 /// IR optimization or other unnecessary stages.
 class LITE_API MobileConfig : public ConfigBase {
+  // whether to load data from memory. Model data will be loaded from memory
+  // buffer if model_from_memory_ is true.
+  bool model_from_memory_{false};
+
+  // model data readed from file or memory buffer in combined format.
+  std::string lite_model_file_;
+
+  // NOTE: This is a deprecated variable and will be removed in latter release.
   std::string model_buffer_;
   std::string param_buffer_;
-  bool model_from_memory_{false};
 
  public:
+  // set model data in combined format, `set_model_from_file` refers to loading
+  // model from file, set_model_from_buffer refers to loading model from memory
+  // buffer
+  void set_model_from_file(const std::string& x);
+  void set_model_from_buffer(const std::string& x);
+  // return model data in lite_model_file_, which is in combined format.
+  const std::string& lite_model_file() const { return lite_model_file_; }
+
+  // return model_from_memory_, which indicates whether to load model from
+  // memory buffer.
+  bool model_from_memory() const { return model_from_memory_; }
+
+  // NOTE: This is a deprecated API and will be removed in latter release.
   void set_model_buffer(const char* model_buffer,
                         size_t model_buffer_size,
                         const char* param_buffer,
-                        size_t param_buffer_size) {
-    model_buffer_ = std::string(model_buffer, model_buffer + model_buffer_size);
-    param_buffer_ = std::string(param_buffer, param_buffer + param_buffer_size);
-    model_from_memory_ = true;
-  }
+                        size_t param_buffer_size);
 
-  bool model_from_memory() const { return model_from_memory_; }
+  // NOTE: This is a deprecated API and will be removed in latter release.
   const std::string& model_buffer() const { return model_buffer_; }
+
+  // NOTE: This is a deprecated API and will be removed in latter release.
   const std::string& param_buffer() const { return param_buffer_; }
 };
 
diff --git a/lite/api/paddle_api_test.cc b/lite/api/paddle_api_test.cc
index 69d544c3decac9f312bc9eb03cdc6c3702c5032b..9213a24e5c0614550a098c4de8d97b6cf6695177 100644
--- a/lite/api/paddle_api_test.cc
+++ b/lite/api/paddle_api_test.cc
@@ -72,7 +72,7 @@ TEST(CxxApi, run) {
 #ifdef LITE_WITH_LIGHT_WEIGHT_FRAMEWORK
 TEST(LightApi, run) {
   lite_api::MobileConfig config;
-  config.set_model_dir(FLAGS_model_dir + ".opt2.naive");
+  config.set_model_from_file(FLAGS_model_dir + ".opt2.naive.nb");
 
   auto predictor = lite_api::CreatePaddlePredictor(config);
 
@@ -109,16 +109,11 @@ TEST(LightApi, run) {
 // Demo2 for Loading model from memory
 TEST(MobileConfig, LoadfromMemory) {
   // Get naive buffer
-  auto model_path = std::string(FLAGS_model_dir) + ".opt2.naive/__model__.nb";
-  auto params_path = std::string(FLAGS_model_dir) + ".opt2.naive/param.nb";
-  std::string model_buffer = lite::ReadFile(model_path);
-  size_t size_model = model_buffer.length();
-  std::string params_buffer = lite::ReadFile(params_path);
-  size_t size_params = params_buffer.length();
+  auto model_file = std::string(FLAGS_model_dir) + ".opt2.naive.nb";
+  std::string model_buffer = lite::ReadFile(model_file);
   // set model buffer and run model
   lite_api::MobileConfig config;
-  config.set_model_buffer(
-      model_buffer.c_str(), size_model, params_buffer.c_str(), size_params);
+  config.set_model_from_buffer(model_buffer);
 
   auto predictor = lite_api::CreatePaddlePredictor(config);
   auto input_tensor = predictor->GetInput(0);
diff --git a/lite/api/python/pybind/pybind.cc b/lite/api/python/pybind/pybind.cc
index 7d4ed4e98701a5328b0f05387dc73ad8b93dfe18..2dfe0c49490ecd13e8a3ce480807bdf3875348b7 100644
--- a/lite/api/python/pybind/pybind.cc
+++ b/lite/api/python/pybind/pybind.cc
@@ -116,6 +116,8 @@ void BindLiteMobileConfig(py::module *m) {
   py::class_<MobileConfig> mobile_config(*m, "MobileConfig");
 
   mobile_config.def(py::init<>())
+      .def("set_model_from_file", &MobileConfig::set_model_from_file)
+      .def("set_model_from_buffer", &MobileConfig::set_model_from_buffer)
       .def("set_model_dir", &MobileConfig::set_model_dir)
       .def("model_dir", &MobileConfig::model_dir)
       .def("set_model_buffer", &MobileConfig::set_model_buffer)
diff --git a/lite/core/mir/subgraph/subgraph_pass_test.cc b/lite/core/mir/subgraph/subgraph_pass_test.cc
index a56c364f975fa6c3f82e1bbbb4489c93eb6ab724..252517939990d8ce48083badb342c22fae1459c6 100644
--- a/lite/core/mir/subgraph/subgraph_pass_test.cc
+++ b/lite/core/mir/subgraph/subgraph_pass_test.cc
@@ -157,7 +157,7 @@ std::shared_ptr<lite_api::PaddlePredictor> TestModel(
                                 lite_api::LiteModelType::kNaiveBuffer);
   // Load optimized model
   lite_api::MobileConfig mobile_config;
-  mobile_config.set_model_dir(optimized_model_dir);
+  mobile_config.set_model_from_file(optimized_model_dir + ".nb");
   mobile_config.set_power_mode(lite_api::PowerMode::LITE_POWER_HIGH);
   mobile_config.set_threads(1);
   predictor = lite_api::CreatePaddlePredictor(mobile_config);
diff --git a/lite/core/version.h.in b/lite/core/version.h.in
index 3082adc5abecb20f5ce19032177fc7cdb75299ff..d34c32073b852a50b5d26984ed4812ac4f38a870 100644
--- a/lite/core/version.h.in
+++ b/lite/core/version.h.in
@@ -42,7 +42,7 @@ static std::string version() {
 
   std::string tag = paddlelite_tag();
   if (tag.empty()) {
-    ss << paddlelite_branch() << "(" << paddlelite_commit() << ")";
+    ss << paddlelite_commit();
   } else {
     ss << tag;
   }
diff --git a/lite/model_parser/model_parser.cc b/lite/model_parser/model_parser.cc
index 0dcb8e1eeab4b07d533a1bfc57cb8d9ca38b4d82..5d00570703f2caaf71ff5b5e1e6c3ad9e27eb6f2 100644
--- a/lite/model_parser/model_parser.cc
+++ b/lite/model_parser/model_parser.cc
@@ -20,6 +20,7 @@
 #include "lite/core/scope.h"
 #include "lite/core/tensor.h"
 #include "lite/core/variable.h"
+#include "lite/core/version.h"
 #include "lite/model_parser/desc_apis.h"
 #include "lite/model_parser/naive_buffer/combined_params_desc.h"
 #include "lite/model_parser/naive_buffer/param_desc.h"
@@ -536,7 +537,7 @@ void SaveCombinedParamsNaive(const std::string &path,
   }
 
   pt_desc.Save();
-  table.SaveToFile(path);
+  table.AppendToFile(path);
 }
 
 void SaveModelNaive(const std::string &model_dir,
@@ -545,30 +546,46 @@ void SaveModelNaive(const std::string &model_dir,
                     bool combined) {
   MkDirRecur(model_dir);
   // Save program
-  const std::string prog_path = model_dir + "/__model__.nb";
+  const std::string prog_path = model_dir + ".nb";
   naive_buffer::BinaryTable table;
   naive_buffer::proto::ProgramDesc nb_proto_prog(&table);
   naive_buffer::ProgramDesc nb_prog(&nb_proto_prog);
   TransformProgramDescCppToAny(cpp_prog, &nb_prog);
   nb_proto_prog.Save();
-  table.SaveToFile(prog_path);
 
+  // Save meta_version(uint16) into file
+  naive_buffer::BinaryTable meta_version_table;
+  meta_version_table.Require(sizeof(uint16_t));
+  uint16_t meta_version = 0;
+  memcpy(meta_version_table.cursor(), &meta_version, sizeof(uint16_t));
+  meta_version_table.Consume(sizeof(uint16_t));
+  meta_version_table.SaveToFile(prog_path);
+
+  // Save lite_version(char[16]) into file
+  const int paddle_version_length = 16 * sizeof(char);
+  naive_buffer::BinaryTable paddle_version_table;
+  paddle_version_table.Require(paddle_version_length);
+  std::string paddle_version = version();
+  memcpy(paddle_version_table.cursor(),
+         paddle_version.c_str(),
+         paddle_version_length);
+  paddle_version_table.Consume(paddle_version_length);
+  paddle_version_table.AppendToFile(prog_path);
+  VLOG(4) << "paddle_version:" << paddle_version << std::endl;
+
+  // Save topology_size(uint64) into file
+  naive_buffer::BinaryTable topology_size_table;
+  topology_size_table.Require(sizeof(uint64_t));
+  uint64_t topology_size = table.size();
+  memcpy(topology_size_table.cursor(), &topology_size, sizeof(uint64_t));
+  topology_size_table.Consume(sizeof(uint64_t));
+  topology_size_table.AppendToFile(prog_path);
+
+  // save topology data into model file
+  table.AppendToFile(prog_path);
   // Save Params
-  // NOTE: Only main block be used now.
-  if (combined) {
-    const std::string combined_params_path = model_dir + "/param.nb";
-    SaveCombinedParamsNaive(combined_params_path, exec_scope, cpp_prog);
-  } else {
-    auto prog = cpp_prog;
-    auto &main_block_desc = *prog.GetBlock<cpp::BlockDesc>(0);
-    for (size_t i = 0; i < main_block_desc.VarsSize(); ++i) {
-      auto &var = *main_block_desc.GetVar<cpp::VarDesc>(i);
-      if (var.Name() == "feed" || var.Name() == "fetch" || !var.Persistable())
-        continue;
-      const std::string path = model_dir + "/" + var.Name() + ".nb";
-      SaveParamNaive(path, exec_scope, var.Name());
-    }
-  }
+  SaveCombinedParamsNaive(prog_path, exec_scope, cpp_prog);
+
   LOG(INFO) << "Save naive buffer model in '" << model_dir << "' successfully";
 }
 #endif
@@ -638,14 +655,15 @@ void LoadParamNaive(const std::string &path,
 }
 
 void LoadCombinedParamsNaive(const std::string &path,
+                             const uint64_t &offset,
                              lite::Scope *scope,
                              const cpp::ProgramDesc &cpp_prog,
                              bool params_from_memory) {
   naive_buffer::BinaryTable table;
   if (params_from_memory) {
-    table.LoadFromMemory(path.c_str(), path.length());
+    table.LoadFromMemory(path.c_str() + offset, path.length() - offset);
   } else {
-    table.LoadFromFile(path);
+    table.LoadFromFile(path, offset, 0);
   }
   naive_buffer::proto::CombinedParamsDesc pt_desc(&table);
   pt_desc.Load();
@@ -693,7 +711,7 @@ void LoadModelNaive(const std::string &model_dir,
   // NOTE: Only main block be used now.
   if (combined) {
     const std::string combined_params_path = model_dir + "/param.nb";
-    LoadCombinedParamsNaive(combined_params_path, scope, *cpp_prog, false);
+    LoadCombinedParamsNaive(combined_params_path, 0, scope, *cpp_prog, false);
   } else {
     auto &prog = *cpp_prog;
     auto &main_block_desc = *prog.GetBlock<cpp::BlockDesc>(0);
@@ -718,6 +736,84 @@ void LoadModelNaive(const std::string &model_dir,
   VLOG(4) << "Load naive buffer model in '" << model_dir << "' successfully";
 }
 
+/*
+ * Binary structure of naive_buffer model: model.nb
+ * ----------------------------------------------------------
+ * |       |    PART         |   Precision |   Length(byte) |
+ * |   1   |  meta_version   |   uint16_t  |       2        |
+ * |   2   |  opt_version    |   char[16]  |      16        |
+ * |   3   |  topo_size      |   uint64_t  |       8        |
+ * |   4   |  topo_data      |   char[]    | topo_size byte |
+ * |   5   |  param_data     |   char[]    |                |
+ * ----------------------------------------------------------
+ *  Meaning of each part:
+ *      meta_version: meata_version, 0 default.
+ *      opt_version:  lite_version of opt tool that transformed this model.
+ *      topo_size:    length of `topo_data`.
+ *      topo_data:    contains model's topology data.
+ *      param_data:   contains model's params data.
+*/
+
+// usage: LoadModelNaiveFromFile is used for loading model from file.
+template <typename T>
+void ReadModelDataFromFile(T *data,
+                           const std::string &prog_path,
+                           uint64_t *offset,
+                           const uint64_t &size) {
+  naive_buffer::BinaryTable data_table;
+  data_table.LoadFromFile(prog_path, *offset, size);
+  memcpy(data, data_table.cursor(), size);
+  *offset = *offset + size;
+}
+
+void LoadModelNaiveFromFile(const std::string &filename,
+                            Scope *scope,
+                            cpp::ProgramDesc *cpp_prog) {
+  CHECK(cpp_prog);
+  CHECK(scope);
+  cpp_prog->ClearBlocks();
+  // ModelFile
+  const std::string prog_path = filename;
+
+  // Offset
+  uint64_t offset = 0;
+
+  // (1)get meta version
+  uint16_t meta_version;
+  ReadModelDataFromFile<uint16_t>(
+      &meta_version, prog_path, &offset, sizeof(uint16_t));
+  VLOG(4) << "Meta_version:" << meta_version;
+
+  // (2)get opt version
+  char opt_version[16];
+  const uint64_t paddle_version_length = 16 * sizeof(char);
+  ReadModelDataFromFile<char>(
+      opt_version, prog_path, &offset, paddle_version_length);
+  VLOG(4) << "Opt_version:" << opt_version;
+
+  // (3)get topo_size
+  uint64_t topo_size;
+  ReadModelDataFromFile<uint64_t>(
+      &topo_size, prog_path, &offset, sizeof(uint64_t));
+
+  // (4)get topo data
+  naive_buffer::BinaryTable topo_table;
+  topo_table.LoadFromFile(prog_path, offset, topo_size);
+  offset = offset + topo_size;
+  // transform topo_data into cpp::ProgramDesc
+  naive_buffer::proto::ProgramDesc nb_proto_prog(&topo_table);
+  nb_proto_prog.Load();
+  naive_buffer::ProgramDesc nb_prog(&nb_proto_prog);
+  TransformProgramDescAnyToCpp(nb_prog, cpp_prog);
+
+  // (5)Load Params
+  LoadCombinedParamsNaive(prog_path, offset, scope, *cpp_prog, false);
+
+  VLOG(4) << "Load naive buffer model in '" << filename << "' successfully";
+}
+
+// warning: this is an old inference and is not suggested.
+// todo: this inference will be abandened in release/v3.0.0
 void LoadModelNaiveFromMemory(const std::string &model_buffer,
                               const std::string &param_buffer,
                               Scope *scope,
@@ -741,7 +837,64 @@ void LoadModelNaiveFromMemory(const std::string &model_buffer,
   // Load Params
   // NOTE: Only main block be used now.
   // only combined Params are supported in Loading Model from memory
-  LoadCombinedParamsNaive(param_buffer, scope, *cpp_prog, true);
+  LoadCombinedParamsNaive(param_buffer, 0, scope, *cpp_prog, true);
+
+  VLOG(4) << "Load model from naive buffer memory successfully";
+}
+
+// usage: LoadModelNaiveFromMemory is used for loading naive model from memory
+template <typename T>
+void ReadModelDataFromBuffer(T *data,
+                             const std::string &model_buffer,
+                             uint64_t *offset,
+                             const uint64_t &size) {
+  naive_buffer::BinaryTable data_table;
+  data_table.LoadFromMemory(model_buffer.c_str() + *offset, size);
+  memcpy(data, data_table.cursor(), size);
+  *offset = *offset + size;
+}
+void LoadModelNaiveFromMemory(const std::string &model_buffer,
+                              Scope *scope,
+                              cpp::ProgramDesc *cpp_prog) {
+  CHECK(cpp_prog);
+  CHECK(scope);
+  cpp_prog->ClearBlocks();
+
+  // Offset
+  uint64_t offset = 0;
+
+  // (1)get meta version
+  uint16_t meta_version;
+  ReadModelDataFromBuffer<uint16_t>(
+      &meta_version, model_buffer, &offset, sizeof(uint16_t));
+  VLOG(4) << "Meta_version:" << meta_version;
+
+  // (2)get opt version
+  char opt_version[16];
+  const uint64_t paddle_version_length = 16 * sizeof(char);
+  ReadModelDataFromBuffer<char>(
+      opt_version, model_buffer, &offset, paddle_version_length);
+  VLOG(4) << "Opt_version:" << opt_version;
+
+  // (3)get topo_size and topo_data
+  uint64_t topo_size;
+  ReadModelDataFromBuffer<uint64_t>(
+      &topo_size, model_buffer, &offset, sizeof(uint64_t));
+  naive_buffer::BinaryTable table;
+  table.LoadFromMemory(model_buffer.c_str() + offset, topo_size);
+  offset = offset + topo_size;
+
+  naive_buffer::proto::ProgramDesc nb_proto_prog(&table);
+  nb_proto_prog.Load();
+  naive_buffer::ProgramDesc nb_prog(&nb_proto_prog);
+
+  // Transform to cpp::ProgramDesc
+  TransformProgramDescAnyToCpp(nb_prog, cpp_prog);
+
+  // Load Params
+  // NOTE: Only main block be used now.
+  // only combined Params are supported in Loading Model from memory
+  LoadCombinedParamsNaive(model_buffer, offset, scope, *cpp_prog, true);
 
   VLOG(4) << "Load model from naive buffer memory successfully";
 }
diff --git a/lite/model_parser/model_parser.h b/lite/model_parser/model_parser.h
index bca7533c24af517994dae677c7b63e088f2ef1ca..e4641f69ada380c91f69280290dd020ea27d2ad1 100644
--- a/lite/model_parser/model_parser.h
+++ b/lite/model_parser/model_parser.h
@@ -94,15 +94,22 @@ void LoadParamNaive(const std::string& path,
                     lite::Scope* scope,
                     const std::string& name);
 
+// warning:this old inference will be abandened in release/v3.0.0
+// and LoadModelNaiveFromFile is suggested.
 void LoadModelNaive(const std::string& model_dir,
                     lite::Scope* scope,
                     cpp::ProgramDesc* prog,
                     bool combined = true);
-
+void LoadModelNaiveFromFile(const std::string& filename,
+                            lite::Scope* scope,
+                            cpp::ProgramDesc* prog);
 void LoadModelNaiveFromMemory(const std::string& model_buffer,
                               const std::string& param_buffer,
                               lite::Scope* scope,
                               cpp::ProgramDesc* cpp_prog);
+void LoadModelNaiveFromMemory(const std::string& model_buffer,
+                              lite::Scope* scope,
+                              cpp::ProgramDesc* cpp_prog);
 
 }  // namespace lite
 }  // namespace paddle
diff --git a/lite/model_parser/model_parser_test.cc b/lite/model_parser/model_parser_test.cc
index 58083027849cc007bce80bd10004d0a13259fda7..d9c0f501c37862236cacd2624dc70c8cf1dacc86 100644
--- a/lite/model_parser/model_parser_test.cc
+++ b/lite/model_parser/model_parser_test.cc
@@ -121,17 +121,23 @@ TEST(ModelParser, SaveModelNaive) {
   SaveModelNaive(save_pb_model_path, scope, prog);
 }
 
+TEST(ModelParser, LoadModelNaiveFromFile) {
+  CHECK(!FLAGS_model_dir.empty());
+  cpp::ProgramDesc prog;
+  Scope scope;
+
+  auto model_path = std::string(FLAGS_model_dir) + ".saved.naive.nb";
+  LoadModelNaiveFromFile(model_path, &scope, &prog);
+}
+
 TEST(ModelParser, LoadModelNaiveFromMemory) {
   CHECK(!FLAGS_model_dir.empty());
   cpp::ProgramDesc prog;
   Scope scope;
 
-  auto model_path = std::string(FLAGS_model_dir) + ".saved.naive/__model__.nb";
-  auto params_path = std::string(FLAGS_model_dir) + ".saved.naive/param.nb";
+  auto model_path = std::string(FLAGS_model_dir) + ".saved.naive.nb";
   std::string model_buffer = lite::ReadFile(model_path);
-  std::string params_buffer = lite::ReadFile(params_path);
-
-  LoadModelNaiveFromMemory(model_buffer, params_buffer, &scope, &prog);
+  LoadModelNaiveFromMemory(model_buffer, &scope, &prog);
 }
 
 }  // namespace lite
diff --git a/lite/model_parser/naive_buffer/naive_buffer.cc b/lite/model_parser/naive_buffer/naive_buffer.cc
index cefaf0c28a34a70c095362e9972c9ef99d5fa80c..02538602fb5b5ae319d1041d501a87c212e47d2d 100644
--- a/lite/model_parser/naive_buffer/naive_buffer.cc
+++ b/lite/model_parser/naive_buffer/naive_buffer.cc
@@ -44,24 +44,37 @@ void BinaryTable::SaveToFile(const std::string &filename) const {
   fclose(fp);
 }
 
-void BinaryTable::LoadFromFile(const std::string &filename) {
-  // get file size
+void BinaryTable::AppendToFile(const std::string &filename) const {
+  FILE *fp = fopen(filename.c_str(), "ab");
+  CHECK(fp) << "Unable to open file: " << filename;
+  if (fwrite(reinterpret_cast<const char *>(data()), 1, size(), fp) != size()) {
+    fclose(fp);
+    LOG(FATAL) << "Write file error: " << filename;
+  }
+  fclose(fp);
+}
+
+void BinaryTable::LoadFromFile(const std::string &filename,
+                               const size_t &offset,
+                               const size_t &size) {
+  // open file in readonly mode
   FILE *fp = fopen(filename.c_str(), "rb");
   CHECK(fp) << "Unable to open file: " << filename;
-  fseek(fp, 0L, SEEK_END);
-  size_t file_size = ftell(fp);
-  LOG(INFO) << "file size " << file_size;
-
-  // load data.
-  fseek(fp, 0L, SEEK_SET);
-  Require(file_size);
-  if (fread(reinterpret_cast<char *>(&bytes_[0]), 1, file_size, fp) !=
-      file_size) {
+  // move fstream pointer backward for size of offset
+  size_t buffer_size = size;
+  if (size == 0) {
+    fseek(fp, 0L, SEEK_END);
+    buffer_size = ftell(fp) - offset;
+  }
+  fseek(fp, offset, SEEK_SET);
+  Require(buffer_size);
+  // read data of `size` into binary_data_variable:`bytes_`
+  if (fread(reinterpret_cast<char *>(&bytes_[0]), 1, buffer_size, fp) !=
+      buffer_size) {
     fclose(fp);
     LOG(FATAL) << "Read file error: " << filename;
   }
   fclose(fp);
-
   // Set readonly.
   is_mutable_mode_ = false;
 }
diff --git a/lite/model_parser/naive_buffer/naive_buffer.h b/lite/model_parser/naive_buffer/naive_buffer.h
index 9be2be954328e757e79a880f34b49c3f0cf77c7a..5be17856a25aabfed81ae88d80e788c8dd2be4bc 100644
--- a/lite/model_parser/naive_buffer/naive_buffer.h
+++ b/lite/model_parser/naive_buffer/naive_buffer.h
@@ -61,8 +61,12 @@ struct BinaryTable {
 
   /// Serialize the table to a binary buffer.
   void SaveToFile(const std::string& filename) const;
+  void AppendToFile(const std::string& filename) const;
 
-  void LoadFromFile(const std::string& filename);
+  //  void LoadFromFile(const std::string& filename);
+  void LoadFromFile(const std::string& filename,
+                    const size_t& offset = 0,
+                    const size_t& size = 0);
   void LoadFromMemory(const char* buffer, size_t buffer_size);
 };
 
diff --git a/lite/tools/build.sh b/lite/tools/build.sh
index 4dfd8f6c70efba9e87811540eb0e5726198cb8f2..7bb330b28bc51ca4a241831bd320cb25474a74cd 100755
--- a/lite/tools/build.sh
+++ b/lite/tools/build.sh
@@ -14,7 +14,7 @@ readonly NUM_PROC=${LITE_BUILD_THREADS:-4}
 
 # global variables
 BUILD_EXTRA=OFF
-BUILD_JAVA=ON
+BUILD_JAVA=OFF
 BUILD_PYTHON=OFF
 BUILD_DIR=$(pwd)
 OPTMODEL_DIR=""
@@ -72,7 +72,7 @@ function build_opt {
       -DWITH_TESTING=OFF \
       -DLITE_BUILD_EXTRA=ON \
       -DWITH_MKL=OFF
-    make OPT -j$NUM_PROC
+    make opt -j$NUM_PROC
 }
 
 function make_tiny_publish_so {
diff --git a/lite/tools/ci_build.sh b/lite/tools/ci_build.sh
index 9b1bc94daa7ea111c05ff7dbf3e575f3093a2eb2..1960dc1e1506f9742cdd9be41d5448c646c026af 100755
--- a/lite/tools/ci_build.sh
+++ b/lite/tools/ci_build.sh
@@ -519,7 +519,7 @@ function test_model_optimize_tool_compile {
     cd $workspace
     cd build
     cmake .. -DWITH_LITE=ON -DLITE_ON_MODEL_OPTIMIZE_TOOL=ON -DWITH_TESTING=OFF -DLITE_BUILD_EXTRA=ON
-    make OPT -j$NUM_CORES_FOR_COMPILE
+    make opt -j$NUM_CORES_FOR_COMPILE
 }
 
 function _test_paddle_code_generator {