diff --git a/paddle/fluid/lite/api/cxx_api_bin.cc b/paddle/fluid/lite/api/cxx_api_bin.cc
new file mode 100644
index 0000000000000000000000000000000000000000..d19a95b34a9adf758cc238ae56103c5782689fa4
--- /dev/null
+++ b/paddle/fluid/lite/api/cxx_api_bin.cc
@@ -0,0 +1,72 @@
+#include "paddle/fluid/lite/api/cxx_api.h"
+#include "paddle/fluid/lite/core/mir/passes.h"
+#include "paddle/fluid/lite/core/op_registry.h"
+
+namespace paddle {
+namespace lite {
+
+void Run(const char* model_dir) {
+  lite::LightPredictor predictor;
+#ifndef LITE_WITH_CUDA
+  std::vector<Place> valid_places({Place{TARGET(kHost), PRECISION(kFloat)}});
+#else
+  std::vector<Place> valid_places({
+      Place{TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW)},
+      Place{TARGET(kCUDA), PRECISION(kFloat), DATALAYOUT(kNCHW)},
+      Place{TARGET(kCUDA), PRECISION(kAny), DATALAYOUT(kNCHW)},
+      Place{TARGET(kHost), PRECISION(kAny), DATALAYOUT(kNCHW)},
+      Place{TARGET(kCUDA), PRECISION(kAny), DATALAYOUT(kAny)},
+      Place{TARGET(kHost), PRECISION(kAny), DATALAYOUT(kAny)},
+  });
+#endif
+
+  predictor.Build(model_dir, Place{TARGET(kCUDA), PRECISION(kFloat)},
+                  valid_places);
+
+  auto* input_tensor = predictor.GetInput(0);
+  input_tensor->Resize(DDim(std::vector<DDim::value_type>({100, 100})));
+  auto* data = input_tensor->mutable_data<float>();
+  for (int i = 0; i < 100 * 100; i++) {
+    data[i] = i;
+  }
+
+  LOG(INFO) << "input " << *input_tensor;
+
+  predictor.Run();
+
+  auto* out = predictor.GetOutput(0);
+  LOG(INFO) << out << " memory size " << out->data_size();
+  LOG(INFO) << "out " << out->data<float>()[0];
+  LOG(INFO) << "out " << out->data<float>()[1];
+  LOG(INFO) << "dims " << out->dims();
+  LOG(INFO) << "out " << *out;
+}
+
+}  // namespace lite
+}  // namespace paddle
+
+int main(int argc, char** argv ) {
+  CHECK_EQ(argc, 2) << "usage: ./cmd <model_dir>";
+  paddle::lite::Run(argv[1]);
+
+  return 0;
+}
+
+
+USE_LITE_OP(mul);
+USE_LITE_OP(fc);
+USE_LITE_OP(scale);
+USE_LITE_OP(feed);
+USE_LITE_OP(fetch);
+USE_LITE_OP(io_copy);
+USE_LITE_KERNEL(fc, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(mul, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(scale, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(feed, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(fetch, kHost, kAny, kAny, def);
+
+#ifdef LITE_WITH_CUDA
+USE_LITE_KERNEL(mul, kCUDA, kFloat, kNCHW, def);
+USE_LITE_KERNEL(io_copy, kCUDA, kAny, kAny, host_to_device);
+USE_LITE_KERNEL(io_copy, kCUDA, kAny, kAny, device_to_host);
+#endif
diff --git a/paddle/fluid/lite/core/framework.proto b/paddle/fluid/lite/core/framework.proto
new file mode 120000
index 0000000000000000000000000000000000000000..d98f7d057fe065712598f0f27a94990d12c7f85a
--- /dev/null
+++ b/paddle/fluid/lite/core/framework.proto
@@ -0,0 +1 @@
+../../../fluid/framework/framework.proto
\ No newline at end of file
diff --git a/paddle/fluid/lite/utils/cp_logging.cc b/paddle/fluid/lite/utils/cp_logging.cc
new file mode 100644
index 0000000000000000000000000000000000000000..d72cf0d782fd8b17516242440995cb9fe5eb389b
--- /dev/null
+++ b/paddle/fluid/lite/utils/cp_logging.cc
@@ -0,0 +1,5 @@
+#include "paddle/fluid/lite/utils/cp_logging.h"
+
+namespace paddle {
+namespace lite {}  // namespace lite
+}  // namespace paddle
diff --git a/paddle/fluid/lite/utils/cp_logging.h b/paddle/fluid/lite/utils/cp_logging.h
new file mode 100644
index 0000000000000000000000000000000000000000..885670cc28bfe287850752bddb7f0c8e98e55d56
--- /dev/null
+++ b/paddle/fluid/lite/utils/cp_logging.h
@@ -0,0 +1,5 @@
+#ifdef LITE_WITH_LIGHT_WEIGHT_FRAMEWORK
+#include "paddle/fluid/lite/utils/logging.h"
+#else  // LITE_WITH_LIGHT_WEIGHT_FRAMEWORK
+#include <glog/logging.h>
+#endif  // LITE_WITH_LIGHT_WEIGHT_FRAMEWORK