diff --git a/libs/ai-endpoints/langchain_nvidia_ai_endpoints/_common.py b/libs/ai-endpoints/langchain_nvidia_ai_endpoints/_common.py
index 2bde648a..454b1d7d 100644
--- a/libs/ai-endpoints/langchain_nvidia_ai_endpoints/_common.py
+++ b/libs/ai-endpoints/langchain_nvidia_ai_endpoints/_common.py
@@ -523,7 +523,7 @@ def get_req_stream(
         }
 
         response = self.get_session_fn().post(
-            **self.__add_authorization(self.last_inputs)
+            stream=True, **self.__add_authorization(self.last_inputs)
         )
         self._try_raise(response)
         call = self.copy()
diff --git a/libs/ai-endpoints/tests/integration_tests/test_streaming.py b/libs/ai-endpoints/tests/integration_tests/test_streaming.py
new file mode 100644
index 00000000..158f81a2
--- /dev/null
+++ b/libs/ai-endpoints/tests/integration_tests/test_streaming.py
@@ -0,0 +1,27 @@
+import time
+
+from langchain_nvidia_ai_endpoints import ChatNVIDIA
+
+
+def test_ttft(chat_model: str, mode: dict) -> None:
+    # we had an issue where streaming took a long time to start. the issue
+    # was all streamed results were collected before yielding them to the
+    # user. this test tries to detect the incorrect behavior.
+    #
+    # warning:
+    #   - this can false positive if the model itself is slow to start
+    #   - this can false nagative if there is a delay after the first chunk
+    #
+    # potential mitigation for false negative is to check mean & stdev and
+    # filter outliers.
+    #
+    # credit to Pouyan Rezakhani for finding this issue
+    llm = ChatNVIDIA(model=chat_model, **mode)
+    chunk_times = [time.time()]
+    for chunk in llm.stream("Count to 1000 by 2s, e.g. 2 4 6 8 ...", max_tokens=512):
+        chunk_times.append(time.time())
+    ttft = chunk_times[1] - chunk_times[0]
+    total_time = chunk_times[-1] - chunk_times[0]
+    assert ttft < (
+        total_time / 2
+    ), "potential streaming issue, TTFT should be less than half of the total time"