diff --git a/modelopt/torch/quantization/extensions.py b/modelopt/torch/quantization/extensions.py
index a65396d64ff..3eac516d287 100644
--- a/modelopt/torch/quantization/extensions.py
+++ b/modelopt/torch/quantization/extensions.py
@@ -15,8 +15,12 @@
 
 """Module to load C++ / CUDA extensions."""
 
+import os
 from pathlib import Path
 
+import torch
+from packaging.version import Version
+
 from modelopt.torch.utils import load_cpp_extension
 
 __all__ = ["get_cuda_ext", "get_cuda_ext_fp8", "get_cuda_ext_mx", "precompile"]
@@ -25,6 +29,18 @@
 kernels_gemm = path.parent / "kernels" / "quantization" / "gemm"
 
 
+def _get_mx_cuda_cflags() -> list[str]:
+    flags = ["--use_fast_math"]
+    if (
+        os.name == "nt"
+        and torch.version.cuda is not None
+        and Version(torch.version.cuda) >= Version("13")
+    ):
+        # CUDA 13 CCCL requires MSVC's standard-conforming preprocessor.
+        flags.extend(["-Xcompiler", "/Zc:preprocessor"])
+    return flags
+
+
 def get_cuda_ext(raise_if_failed: bool = False):
     """Returns the cuda extension for tensor_quant."""
     if not hasattr(get_cuda_ext, "extension"):
@@ -66,7 +82,7 @@ def get_cuda_ext_mx(raise_if_failed: bool = False):
                 "CUDA extension for MX quantization could not be built and loaded, MX simulated"
                 " quantization will not be available."
             ),
-            extra_cuda_cflags=["--use_fast_math"],
+            extra_cuda_cflags=_get_mx_cuda_cflags(),
             raise_if_failed=raise_if_failed,
         )
     return get_cuda_ext_mx.extension  # type:ignore[attr-defined]
diff --git a/pyproject.toml b/pyproject.toml
index f23ce722e5c..be6ee364e3c 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -55,6 +55,7 @@ dependencies = [
 onnx = [
     "cppimport",
     "cupy-cuda12x; platform_machine != 'aarch64' and platform_machine != 'ARM64' and platform_system != 'Darwin'",
+    "cupy-cuda13x>=14.2.0; python_version > '3.10' and platform_system == 'Windows' and platform_machine == 'ARM64'",
     "lief",
     "ml_dtypes",
     "onnx-graphsurgeon>=0.6.1",
@@ -65,7 +66,7 @@ onnx = [
     # ORT for Windows x64
     "onnxruntime-gpu==1.22.0; platform_system == 'Windows' and platform_machine == 'AMD64'",
     # ORT host and standalone TensorRT-RTX ABI EP for native Windows ARM64.
-    "onnxruntime~=1.24.2; python_version > '3.10' and platform_system == 'Windows' and platform_machine == 'ARM64'",
+    "onnxruntime~=1.26.0; python_version > '3.10' and platform_system == 'Windows' and platform_machine == 'ARM64'",
     "onnxruntime-ep-nv-tensorrt-rtx-cu13==0.4.0; python_version > '3.10' and platform_system == 'Windows' and platform_machine == 'ARM64'",
     # ORT with Python <= 3.10 on supported platforms
     "onnxruntime~=1.22.0; python_version <= '3.10' and (platform_machine == 'aarch64' or platform_system == 'Darwin')",