diff --git a/T/Torch/build_tarballs.jl b/T/Torch/build_tarballs.jl
index 0e92853ae27..62bd91cddba 100644
--- a/T/Torch/build_tarballs.jl
+++ b/T/Torch/build_tarballs.jl
@@ -8,30 +8,47 @@ version = v"1.4.0"
 # Collection of sources required to complete build
 sources = [
     GitSource("https://github.com/dhairyagandhi96/Torch.jl.git", "85bd08d39e7fba29ec4a643f60dd006ed8be8ede"),
-    ArchiveSource("https://download.pytorch.org/libtorch/cu101/libtorch-cxx11-abi-shared-with-deps-1.4.0.zip", "f214bfde532877aa5d4e0803e51a28fa8edd97b6a44b6615f75a70352b6b542e"),
-    ArchiveSource("https://github.com/JuliaGPU/CUDABuilder/releases/download/v0.3.0/CUDNN+CUDA10.1.v7.6.5.x86_64-linux-gnu.tar.gz", "79de5b5085a33bc144b87028e998a1d295a15c3424d6d45b25defe500f616974", unpack_target = "cudnn"),
+    ArchiveSource("https://download.pytorch.org/libtorch/cu101/libtorch-cxx11-abi-shared-with-deps-1.4.0.zip", "f214bfde532877aa5d4e0803e51a28fa8edd97b6a44b6615f75a70352b6b542e", unpack_target="x86_64-linux-gnu"),
+    ArchiveSource("https://download.pytorch.org/libtorch/cu101/libtorch-win-shared-with-deps-1.4.0.zip", "a18a2fa0d952b56a6b9ddc7aae375847eebb5f1bc02fb24f0c4a1f132f8e45c8", unpack_target="x86_64-w64-mingw32"),
+    ArchiveSource("https://github.com/JuliaGPU/CUDABuilder/releases/download/v0.3.0/CUDNN+CUDA10.1.v7.6.5.x86_64-linux-gnu.tar.gz", "79de5b5085a33bc144b87028e998a1d295a15c3424d6d45b25defe500f616974", unpack_target = "x86_64-linux-gnu/cudnn"),
+    ArchiveSource("https://github.com/JuliaGPU/CUDABuilder/releases/download/v0.3.0/CUDNN+CUDA10.1.v7.6.5.x86_64-w64-mingw32.tar.gz", "35f363250ca285315b987561d18871d71ad0403ad92d5259437d5638c3c16d03", unpack_target = "x86_64-w64-mingw32/cudnn"),
 ]
 
 # Bash recipe for building across all platforms
 script = raw"""
 cd $WORKSPACE/srcdir
 
-mv cudnn $prefix
-mv libtorch/share/* $prefix/share/
-mv libtorch/lib/* $prefix/lib/
-rm -r libtorch/lib
-rm -r libtorch/share
-mv libtorch/* $prefix
-rm -r libtorch
+mv $target/cudnn $prefix
+
+if [[ ${target} == x86_64-w64-mingw32 ]]; then
+    mkdir $prefix/lib
+    mv $target/libtorch/bin/* $prefix/bin/
+    rm -r $target/libtorch/bin
+
+    EXTRA_CMAKE_FLAGS="-DCMAKE_SYSTEM_NAME=Windows"
+fi
+
+mv $target/libtorch/share/* $prefix/share/
+mv $target/libtorch/lib/* $prefix/lib/
+rm -r $target/libtorch/lib
+rm -r $target/libtorch/share
+mv $target/libtorch/* $prefix
+rm -r $target/libtorch
 
 mkdir -p /usr/local/cuda/lib64
 cd /usr/local/cuda/lib64
-ln -s ${prefix}/cuda/lib64/libcudart.so libcudart.so
-ln -s ${prefix}/cuda/lib64/libnvToolsExt.so libnvToolsExt.so
+ln -s ${prefix}/cuda/lib64/libcudart.$dlext libcudart.$dlext
+ln -s ${prefix}/cuda/lib64/libnvToolsExt.$dlext libnvToolsExt.$dlext
 
 cd $WORKSPACE/srcdir/Torch.jl/build
+if [[ ${target} == x86_64-w64-mingw32 ]]; then
+    # FindCUDNN.cmake in Torch.jl is very Linux oriented. Patch it up
+    # a little to work in this environment.
+    sed -i 's/libcudnn.so/cudnn64_7.dll/g' FindCUDNN.cmake
+    sed -i 's/lib"/bin"/g' FindCUDNN.cmake
+fi
 mkdir build && cd build
-cmake -DCMAKE_PREFIX_PATH=$prefix -DTorch_DIR=$prefix/share/cmake/Torch -DCUDA_TOOLKIT_ROOT_DIR=$prefix/cuda ..
+cmake -DCMAKE_PREFIX_PATH=$prefix $EXTRA_CMAKE_FLAGS -DTorch_DIR=$prefix/share/cmake/Torch -DCUDA_TOOLKIT_ROOT_DIR=$prefix/cuda ..
 cmake --build .
 
 mkdir -p "${libdir}"
@@ -43,7 +60,9 @@ install_license ${WORKSPACE}/srcdir/Torch.jl/LICENSE
 # These are the platforms we will build for by default, unless further
 # platforms are passed in on the command line
 platforms = [
-    Linux(:x86_64, libc=:glibc, compiler_abi = CompilerABI(cxxstring_abi = :cxx11)),
+    # Temporarily disable Linux build for faster Windows turnaround.
+    #Linux(:x86_64, libc=:glibc, compiler_abi = CompilerABI(cxxstring_abi = :cxx11)),
+    Windows(:x86_64)
 ]
 
 # The products that we will ensure are always built