diff --git a/onnxruntime/core/optimizer/transpose_optimization/onnx_transpose_optimization.cc b/onnxruntime/core/optimizer/transpose_optimization/onnx_transpose_optimization.cc
index 467d0c090070f..8529ccf169f56 100755
--- a/onnxruntime/core/optimizer/transpose_optimization/onnx_transpose_optimization.cc
+++ b/onnxruntime/core/optimizer/transpose_optimization/onnx_transpose_optimization.cc
@@ -454,6 +454,23 @@ static std::optional<DQToLookPast> GetDQWithConstInitializerInputAndSingleConsum
   return result;
 }
 
+// Element types that ONNX QuantizeLinear can produce as its output.
+static bool IsQuantizeLinearOutputType(api::DataType dtype) {
+  switch (dtype) {
+    case api::DataType::INT8:
+    case api::DataType::UINT8:
+    case api::DataType::INT16:
+    case api::DataType::UINT16:
+    case api::DataType::FLOAT8E4M3FN:
+    case api::DataType::FLOAT8E4M3FNUZ:
+    case api::DataType::FLOAT8E5M2:
+    case api::DataType::FLOAT8E5M2FNUZ:
+      return true;
+    default:
+      return false;
+  }
+}
+
 /// <summary>
 /// Insert a Q -> DQ pair after the node following the DQ by using scale and zp info from the preceding DQ node.
 /// DQ -> next node => DQ -> next node -> Q -> DQ.
@@ -528,9 +545,26 @@ static bool MakeQDQNodeUnit(api::GraphRef& graph, const api::NodeRef& dq_node) {
     inputs.push_back(zp_input.value());
   }
 
+  // A zero-point-less DQ with a non-uint8 type needs the new Q's output_dtype pinned, or Q type
+  // inference defaults to uint8 and clashes with the int8 value-info copied below. output_dtype is
+  // ONNX opset 21+ only; if it can't be expressed (older opset, non-ONNX domain, or a type that
+  // QuantizeLinear can't output), skip the push-through so the graph stays valid.
+  std::optional<int64_t> q_output_dtype;
+  if (!zp_input.has_value()) {
+    const api::DataType dq_input_dtype = graph.GetValueInfo(dq_inputs[0])->DType();
+    if (dq_input_dtype != api::DataType::UNDEFINED && dq_input_dtype != api::DataType::UINT8) {
+      const std::optional<int64_t> domain_opset = graph.Opset(dq_domain);
+      if (!IsOnnxDomain(dq_domain) || !domain_opset || *domain_opset < 21 ||
+          !IsQuantizeLinearOutputType(dq_input_dtype)) {
+        return false;
+      }
+      q_output_dtype = static_cast<int64_t>(dq_input_dtype);
+    }
+  }
+
   // Add Q
   auto new_q_node = MakeQuantizeOp(graph, dq_domain, inputs, axis, dq_node.GetAttributeInt("block_size"),
-                                   dq_node.GetAttributeInt("output_dtype"), dq_node.GetAttributeInt("saturate"));
+                                   q_output_dtype, dq_node.GetAttributeInt("saturate"));
   new_q_node->SetLayeringAnnotation(dq_node.GetLayeringAnnotation());
   auto q_node_outputs = new_q_node->Outputs();
 
diff --git a/onnxruntime/test/optimizer/transpose_optimizer_test.cc b/onnxruntime/test/optimizer/transpose_optimizer_test.cc
index 080c382db5d93..2640408df7ab7 100644
--- a/onnxruntime/test/optimizer/transpose_optimizer_test.cc
+++ b/onnxruntime/test/optimizer/transpose_optimizer_test.cc
@@ -3773,6 +3773,37 @@ TEST(TransposeOptimizerTests, TestDequantizeLinearNoAxis) {
 #endif
 }
 
+// Regression test for #28716: pushing a Transpose through a zero-point-less int8 DequantizeLinear
+// inserts a QuantizeLinear that must set output_dtype, else it defaults to uint8 and Resolve() fails.
+TEST(TransposeOptimizerTests, TestDequantizeLinearNoZeroPoint) {
+  auto build_test_case = [&](ModelTestBuilder& builder) {
+    auto* input0_arg = MakeInput<int8_t>(builder, {{2, -1, 6, 3}}, {2, 4, 6, 3}, -128, 127);
+    auto* scale_arg = MakeInput<float>(builder, std::vector<int64_t>{}, std::vector<int64_t>{}, {0.05f});
+    auto* transpose_1_out_0 = builder.MakeIntermediate();
+    auto* dq_out_0 = builder.MakeIntermediate();
+    auto* transpose_2_out_0 = builder.MakeOutput();
+
+    auto& transpose_1 = builder.AddNode("Transpose", {input0_arg}, {transpose_1_out_0});
+    transpose_1.AddAttribute("perm", std::vector<int64_t>{0, 3, 1, 2});
+    builder.AddNode("DequantizeLinear", {transpose_1_out_0, scale_arg}, {dq_out_0});  // no zero-point
+    auto& transpose_2 = builder.AddNode("Transpose", {dq_out_0}, {transpose_2_out_0});
+    transpose_2.AddAttribute("perm", std::vector<int64_t>{0, 2, 3, 1});
+  };
+
+  // opset 21: output_dtype pins the inserted Q's type and the transposes cancel.
+  auto check_cancelled = [](InferenceSessionWrapper& session) {
+    EXPECT_EQ(EstimateTransposeCost(session.GetGraph()), 0);
+  };
+  TransformerTester(build_test_case, check_cancelled, TransformerLevel::Default,
+                    TransformerLevel::Level1, /*opset_version*/ 21);
+
+  // Pre-opset-21 has no output_dtype, so the optimizer must skip the push-through rather than emit
+  // an invalid QuantizeLinear; the model must still initialize (no type-inference crash).
+  auto check_valid = [](InferenceSessionWrapper& /*session*/) {};
+  TransformerTester(build_test_case, check_valid, TransformerLevel::Default,
+                    TransformerLevel::Level1, /*opset_version*/ 19);
+}
+
 TEST(TransposeOptimizerTests, TestCast) {
   auto build_test_case_1 = [&](ModelTestBuilder& builder) {
     auto* input0_arg = MakeInput<int32_t>(builder, {{-1, 4, -1, 5}}, {2, 4, 6, 5}, -1, 5);