diff --git a/tensorflow/lite/tools/optimize/BUILD b/tensorflow/lite/tools/optimize/BUILD
index 185d1d69ce0..b7487b33fae 100644
--- a/tensorflow/lite/tools/optimize/BUILD
+++ b/tensorflow/lite/tools/optimize/BUILD
@@ -181,7 +181,7 @@ tf_cc_test(
         "--test_model_file=$(location //tensorflow/lite/tools/optimize:testdata/single_conv_weights_min_0_max_plus_10.bin)",
     ],
     data = [
-        "//tensorflow/lite/tools/optimize:testdata/single_conv_weights_min_0_max_plus_10.bin",
+        ":testdata/single_conv_weights_min_0_max_plus_10.bin",
     ],
     tags = [
         "tflite_not_portable_android",
@@ -196,6 +196,7 @@ tf_cc_test(
         "//tensorflow/lite/schema:schema_fbs",
         "//tensorflow/lite/schema:schema_utils",
         "//tensorflow/lite/testing:util",
+        "//third_party/eigen3",
         "@com_google_absl//absl/memory",
         "@com_google_absl//absl/strings",
         "@com_google_googletest//:gtest",
diff --git a/tensorflow/lite/tools/optimize/quantization_utils.cc b/tensorflow/lite/tools/optimize/quantization_utils.cc
index 81110071dc9..d2e04338375 100644
--- a/tensorflow/lite/tools/optimize/quantization_utils.cc
+++ b/tensorflow/lite/tools/optimize/quantization_utils.cc
@@ -502,9 +502,14 @@ TfLiteStatus QuantizeTensorFloat16(ModelT* model, TensorT* tensor) {
   // Transform float data to float16.
   std::vector<Eigen::half> quantized_buffer;
   quantized_buffer.resize(num_elements);
-  std::transform(
-      float_vector.begin(), float_vector.end(), quantized_buffer.begin(),
-      [](float a) { return Eigen::half_impl::float_to_half_rtne(a); });
+  constexpr float kMaxFloat16Value = 65504.f;
+  constexpr float kMinFloat16Value = -65504.f;
+  std::transform(float_vector.begin(), float_vector.end(),
+                 quantized_buffer.begin(), [=](float a) {
+                   float clamped = std::min(std::max(a, kMinFloat16Value),
+                                            kMaxFloat16Value);
+                   return Eigen::half_impl::float_to_half_rtne(clamped);
+                 });
 
   char* half_buffer = reinterpret_cast<char*>(quantized_buffer.data());
   model->buffers[tensor->buffer]->data.assign(
diff --git a/tensorflow/lite/tools/optimize/quantization_utils_test.cc b/tensorflow/lite/tools/optimize/quantization_utils_test.cc
index 4ce0d01fd12..76648eaafac 100644
--- a/tensorflow/lite/tools/optimize/quantization_utils_test.cc
+++ b/tensorflow/lite/tools/optimize/quantization_utils_test.cc
@@ -575,6 +575,42 @@ TEST_F(QuantizationUtilsTest, SymmetricQuantizeTensor) {
   EXPECT_EQ(quant_buffer_size * 4, float_buffer_size);
 }
 
+TEST_F(QuantizationUtilsTest, QuantizeFloat16Clamp) {
+  // Create data.
+  auto model = absl::make_unique<ModelT>();
+  auto subgraph = absl::make_unique<tflite::SubGraphT>();
+  auto tensor = absl::make_unique<TensorT>();
+  auto buffer = absl::make_unique<tflite::BufferT>();
+  constexpr int kNumElements = 6;
+  const std::vector<float> weights = {2.0, 1.0, 65504., 65505, -65504., -99999};
+  auto weights_reinterpreted_data =
+      reinterpret_cast<const unsigned char*>(weights.data());
+  buffer->data.assign(weights_reinterpreted_data,
+                      weights_reinterpreted_data + weights.size() * 4);
+  tensor->buffer = 0;
+  tensor->shape = {1, kNumElements};
+
+  // Wire the model.
+  model->subgraphs.push_back(std::move(subgraph));
+  model->subgraphs[0]->tensors.push_back(std::move(tensor));
+  model->buffers.push_back(std::move(buffer));
+
+  // Call and verify.
+  EXPECT_EQ(
+      QuantizeTensorFloat16(model.get(), model->subgraphs[0]->tensors[0].get()),
+      kTfLiteOk);
+  auto weightsf16 = reinterpret_cast<Eigen::half*>(
+      model->buffers[model->subgraphs[0]->tensors[0]->buffer]->data.data());
+  std::vector<float> wf32(kNumElements);
+  std::transform(weightsf16, weightsf16 + 6, wf32.begin(), [](Eigen::half a) {
+    return Eigen::half_impl::half_to_float(a);
+  });
+
+  EXPECT_THAT(wf32,
+              ElementsAreArray({2.0, 1.0, 65504., 65504., -65504., -65504.}));
+  EXPECT_EQ(model->subgraphs[0]->tensors[0]->type, TensorType_FLOAT16);
+}
+
 TEST_F(QuantizationUtilsTest, QuantizeFloat16) {
   // Conv model has weights between 0 and 10.
   // Quantize the weights tensor.