diff --git a/CHANGELOG.rst b/CHANGELOG.rst
index d6397b993a0..757215cd2e3 100755
--- a/CHANGELOG.rst
+++ b/CHANGELOG.rst
@@ -192,6 +192,7 @@ Changelog
 
 **Bug Fixes**
 
+- Fix corrupted ONNX external tensor data when exporting models on Windows by upgrading ONNX. Regenerate affected models from clean weights.
 - Fix ``import modelopt.torch`` failing with ``ModuleNotFoundError`` for ``requests`` or ``huggingface_hub`` when ``nvidia-modelopt`` is installed without the ``[hf]`` extra.
 - Harden the distributed object broadcast/allgather helpers and the Megatron expert-parallel export gather so a compromised peer rank can no longer execute code through a crafted pickle payload: payloads load with ``weights_only=True`` and the export gathers raw bytes instead of pickled objects.
 - Fix LSQ clamping small, representable FP8 block scales too aggressively, which could zero quantized values and suppress scale gradients.
diff --git a/docs/source/getting_started/windows/_installation_windows_arm64.rst b/docs/source/getting_started/windows/_installation_windows_arm64.rst
index 2bafcf0be4e..e50f039a303 100644
--- a/docs/source/getting_started/windows/_installation_windows_arm64.rst
+++ b/docs/source/getting_started/windows/_installation_windows_arm64.rst
@@ -101,7 +101,7 @@ explicit dependency installation avoids selecting the Windows x64 ``onnxruntime-
 
    & $PythonExe -m pip install `
        cppimport cupy-cuda13x lief ml_dtypes `
-       "onnx~=1.21.0" "onnx-graphsurgeon>=0.6.1" `
+       "onnx~=1.23.2" "onnx-graphsurgeon>=0.6.1" `
        "onnxconverter-common~=1.16.0" onnxscript `
        "onnxslim>=0.1.76" "polygraphy>=0.49.22" `
        "onnxruntime>=1.24.2" onnxruntime-ep-nv-tensorrt-rtx-cu13
diff --git a/pyproject.toml b/pyproject.toml
index e585e7fb378..140e08d5ae0 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -59,7 +59,8 @@ onnx = [
     "lief",
     "ml_dtypes",
     "onnx-graphsurgeon>=0.6.1",
-    "onnx~=1.21.0",
+    "onnx~=1.21.0; platform_system != 'Windows'",
+    "onnx~=1.23.2; platform_system == 'Windows'",
     "onnxconverter-common~=1.16.0",
     # Python 3.10 support for ModelOpt ONNX is deprecated and will be removed in a future release.
     # Native Windows ARM64 does not support ModelOpt ONNX on Python 3.10.
diff --git a/tests/unit/onnx/test_external_data.py b/tests/unit/onnx/test_external_data.py
new file mode 100644
index 00000000000..ddb77e997f8
--- /dev/null
+++ b/tests/unit/onnx/test_external_data.py
@@ -0,0 +1,42 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+import onnx
+import pytest
+
+from modelopt.onnx.utils import save_onnx
+
+
+@pytest.mark.parametrize("last_byte", [0x00, 0x0A, 0x1A])
+def test_save_external_data_preserves_tensor_boundaries(tmp_path, last_byte):
+    # Windows text-mode file opening can truncate a trailing Ctrl-Z before append.
+    first = np.full(1024, 0x41, dtype=np.uint8)
+    first[-1] = last_byte
+    second = np.full(1024, 0x42, dtype=np.uint8)
+    tensors = [
+        onnx.numpy_helper.from_array(value, name)
+        for name, value in [("first", first), ("second", second)]
+    ]
+    graph = onnx.helper.make_graph([], "external_data", [], [], initializer=tensors)
+    model = onnx.helper.make_model(graph)
+    path = tmp_path / "model.onnx"
+
+    save_onnx(model, str(path), save_as_external_data=True)
+
+    assert (tmp_path / "model.onnx_data").read_bytes() == first.tobytes() + second.tobytes()
+    restored = onnx.load(path)
+    for actual, expected in zip(restored.graph.initializer, [first, second], strict=True):
+        np.testing.assert_array_equal(onnx.numpy_helper.to_array(actual), expected)