diff --git a/modules/dnn/src/onnx/onnx_graph_simplifier.hpp b/modules/dnn/src/onnx/onnx_graph_simplifier.hpp
index a338e13796..ac2bb295a9 100644
--- a/modules/dnn/src/onnx/onnx_graph_simplifier.hpp
+++ b/modules/dnn/src/onnx/onnx_graph_simplifier.hpp
@@ -42,6 +42,15 @@ Mat getMatFromTensor(const opencv_onnx::TensorProto& tensor_proto, bool uint8ToI
 // or returns -1 when the data_type has no OpenCV equivalent.
 int dataType2cv(int dt);
 
+/** @brief frees a tensor's raw payload once it has been copied into a Mat.
+ *  clear_raw_data() only empties the string and keeps its capacity, so it must be released.
+ */
+inline void releaseONNXTensor(opencv_onnx::TensorProto& tensor_proto)
+{
+    if (!tensor_proto.raw_data().empty())
+        delete tensor_proto.release_raw_data();
+}
+
 CV__DNN_INLINE_NS_END
 }}  // namespace dnn, namespace cv
 
diff --git a/modules/dnn/src/onnx/onnx_importer.cpp b/modules/dnn/src/onnx/onnx_importer.cpp
index e009a3b30b..3720230aa8 100644
--- a/modules/dnn/src/onnx/onnx_importer.cpp
+++ b/modules/dnn/src/onnx/onnx_importer.cpp
@@ -137,13 +137,6 @@ static int onnxDataTypeToCvDepth(int onnxType)
     }
 }
 
-static void releaseONNXTensor(opencv_onnx::TensorProto& tensor_proto)
-{
-    if (!tensor_proto.raw_data().empty()) {
-        delete tensor_proto.release_raw_data();
-    }
-}
-
 Mat readTensorFromONNX(const String& path)
 {
     std::fstream input(path.c_str(), std::ios::in | std::ios::binary);
diff --git a/modules/dnn/src/onnx/onnx_importer2.cpp b/modules/dnn/src/onnx/onnx_importer2.cpp
index 883e2275be..0d21b7390a 100644
--- a/modules/dnn/src/onnx/onnx_importer2.cpp
+++ b/modules/dnn/src/onnx/onnx_importer2.cpp
@@ -534,7 +534,7 @@ LayerParams ONNXImporter2::getLayerParams(const opencv_onnx::NodeProto& node_pro
             }
             else if (attribute_proto.has_t())
             {
-                opencv_onnx::TensorProto tensor = attribute_proto.t();
+                const opencv_onnx::TensorProto& tensor = attribute_proto.t();
                 Mat blob = parseTensor(tensor);
                 lp.blobs.push_back(blob);
                 lp.set("original_dims_of_mat", tensor.dims_size());
@@ -888,9 +888,10 @@ Ptr<Graph> ONNXImporter2::parseGraph(opencv_onnx::GraphProto* graph_proto, bool
     // parse constant tensors
     int n_consts = graph_proto->initializer_size();
     for (int i = 0; i < n_consts; i++) {
-        const opencv_onnx::TensorProto& const_i = graph_proto->initializer(i);
-        Mat t = parseTensor(const_i);
-        netimpl->newConstArg(remap(const_i.name()), t);
+        opencv_onnx::TensorProto* const_i = graph_proto->mutable_initializer(i);
+        Mat t = parseTensor(*const_i);
+        netimpl->newConstArg(remap(const_i->name()), t);
+        releaseONNXTensor(*const_i);
     }
 
     // parse graph inputs
diff --git a/modules/dnn/test/test_backends.cpp b/modules/dnn/test/test_backends.cpp
index 238374cad5..9f94206eea 100644
--- a/modules/dnn/test/test_backends.cpp
+++ b/modules/dnn/test/test_backends.cpp
@@ -37,17 +37,35 @@ public:
         if (!proto.empty())
             proto = findDataFile(proto);
 
-        // Create two networks - with default backend and target and a tested one.
-        Net netDefault = readNet(weights, proto);
-        netDefault.setPreferableBackend(DNN_BACKEND_OPENCV);
-        netDefault.setInput(inp);
+        Mat inp2 = inp.clone();
+        float* inpData = (float*)inp2.data;
+        for (int i = 0; i < inp2.size[0] * inp2.size[1]; ++i)
+        {
+            Mat slice(inp2.size[2], inp2.size[3], CV_32F, inpData);
+            cv::flip(slice, slice, 1);
+            inpData += slice.total();
+        }
 
-        // BUG: https://github.com/opencv/opencv/issues/26349
-        Mat outDefault;
-        if(netDefault.getMainGraph())
-            outDefault = netDefault.forward().clone();
-        else
-            outDefault = netDefault.forward(outputLayer).clone();
+        // Both reference passes run before the tested net is loaded, so netDefault can be
+        // released first. Keeping both alive doubles peak memory on large models.
+        Mat outDefault1, outDefault2;
+        {
+            Net netDefault = readNet(weights, proto);
+            netDefault.setPreferableBackend(DNN_BACKEND_OPENCV);
+
+            // BUG: https://github.com/opencv/opencv/issues/26349
+            netDefault.setInput(inp);
+            if(netDefault.getMainGraph())
+                outDefault1 = netDefault.forward().clone();
+            else
+                outDefault1 = netDefault.forward(outputLayer).clone();
+
+            netDefault.setInput(inp2);
+            if(netDefault.getMainGraph())
+                outDefault2 = netDefault.forward().clone();
+            else
+                outDefault2 = netDefault.forward(outputLayer).clone();
+        }
 
         net = readNet(weights, proto);
         net.setInput(inp);
@@ -64,30 +82,15 @@ public:
         else
             out = net.forward(outputLayer).clone();
 
-        check(outDefault, out, outputLayer, l1, lInf, detectionConfThresh, "First run");
-
-        // Test 2: change input.
-        float* inpData = (float*)inp.data;
-        for (int i = 0; i < inp.size[0] * inp.size[1]; ++i)
-        {
-            Mat slice(inp.size[2], inp.size[3], CV_32F, inpData);
-            cv::flip(slice, slice, 1);
-            inpData += slice.total();
-        }
-        netDefault.setInput(inp);
-        net.setInput(inp);
-
-        if(netDefault.getMainGraph())
-            outDefault = netDefault.forward().clone();
-        else
-            outDefault = netDefault.forward(outputLayer).clone();
+        check(outDefault1, out, outputLayer, l1, lInf, detectionConfThresh, "First run");
 
+        net.setInput(inp2);
         if(net.getMainGraph())
             out = net.forward().clone();
         else
             out = net.forward(outputLayer).clone();
 
-        check(outDefault, out, outputLayer, l1, lInf, detectionConfThresh, "Second run");
+        check(outDefault2, out, outputLayer, l1, lInf, detectionConfThresh, "Second run");
     }
 
     void check(Mat& ref, Mat& out, const std::string& outputLayer, double l1, double lInf,
@@ -124,12 +127,6 @@ TEST_P(DNNTestNetwork, YOLOv8n) {
 TEST_P(DNNTestNetwork, AlexNet)
 {
     applyTestTag(CV_TEST_TAG_MEMORY_1GB);
-    // fc6 needs one contiguous 9216*4096*4 = 144 MiB block, which does not fit
-    // reliably in the 2 GB user address space of a 32-bit process on Windows.
-#ifdef _WIN32
-    if (sizeof(void*) == 4)
-        throw SkipTestException("Skip memory-heavy network on 32-bit (x86) Windows platform");
-#endif
     processNet("dnn/onnx/models/alexnet.onnx", "", Size(227, 227));
     expectNoFallbacksFromIE(net);
     expectNoFallbacksFromCUDA(net);