diff --git a/pyproject.toml b/pyproject.toml
index 774cd08fa5..e295328dea 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -187,7 +187,12 @@ pin_tensorflow_gpu = [
 pin_pytorch_cpu = [
   # https://github.com/pytorch/pytorch/issues/114602
   # macos x86 has been deprecated
-  "torch==2.11.0; platform_machine!='x86_64' or platform_system != 'Darwin'",
+  # Keep the established Python 3.10 pin; use 2.13 for newer CPU tests.
+  # torch 2.13 includes the AOTInductor int_array cache fix for bound methods:
+  # https://github.com/pytorch/pytorch/pull/178147
+  # vesin-torch 0.6.2 requires torch<2.15.
+  "torch==2.11.0; python_version<'3.11' and (platform_machine!='x86_64' or platform_system != 'Darwin')",
+  "torch==2.13.0; python_version>='3.11' and (platform_machine!='x86_64' or platform_system != 'Darwin')",
   "torch; platform_machine=='x86_64' and platform_system == 'Darwin'",
 ]
 pin_pytorch_gpu = [
diff --git a/source/CMakeLists.txt b/source/CMakeLists.txt
index 4ccc18542c..f28545c309 100644
--- a/source/CMakeLists.txt
+++ b/source/CMakeLists.txt
@@ -404,8 +404,10 @@ if(ENABLE_PYTORCH AND NOT DEEPMD_C_ROOT)
   else()
     set(DEEPMD_TORCH_HAS_CUDA FALSE)
   endif()
-  # PyTorch 2.14 raised the minimum C++ standard from C++17 to C++20.
-  if(NOT Torch_VERSION VERSION_LESS "2.14.0")
+  # PyTorch 2.14 requires C++20. MSVC also needs it for the designated
+  # initializers and bit-field defaults already present in PyTorch 2.13 headers.
+  if(NOT Torch_VERSION VERSION_LESS "2.14.0"
+     OR (MSVC AND NOT Torch_VERSION VERSION_LESS "2.13.0"))
     set_if_higher(CMAKE_CXX_STANDARD 20)
   elseif(NOT Torch_VERSION VERSION_LESS "2.1.0")
     set_if_higher(CMAKE_CXX_STANDARD 17)