diff --git a/python/tvm/relay/frontend/tflite.py b/python/tvm/relay/frontend/tflite.py
index 4d607e46c97f..8a81f47f8bcc 100644
--- a/python/tvm/relay/frontend/tflite.py
+++ b/python/tvm/relay/frontend/tflite.py
@@ -775,7 +775,7 @@ def convert_softmax(self, op):
         assert len(output_tensors) == 1, "output tensors length should be 1"
         output_tensor = output_tensors[0]
 
-        params = {"axis": 1}  # 1 is channel
+        params = {"axis": -1}  # -1 is channel
         in_expr = self.get_expr(input_tensor_idx)
 
         # TODO - Naive softmax int8 implementation leads to bad accuracy. Currently, we can
diff --git a/tests/python/frontend/tflite/test_forward.py b/tests/python/frontend/tflite/test_forward.py
index f2941030f0ab..535459915ca3 100644
--- a/tests/python/frontend/tflite/test_forward.py
+++ b/tests/python/frontend/tflite/test_forward.py
@@ -3266,6 +3266,7 @@ def _test_softmax(data):
 def test_forward_softmax():
     """Softmax"""
     _test_softmax(np.arange(6.0, dtype=np.float32).reshape((1, 6)))
+    _test_softmax(np.arange(6.0, dtype=np.float32).reshape((1, 2, 3)))
 
 
 ######################################################################