diff --git a/.github/workflows/compile.yml b/.github/workflows/compile.yml
index 51691dfc4..1f42b1b41 100644
--- a/.github/workflows/compile.yml
+++ b/.github/workflows/compile.yml
@@ -405,7 +405,7 @@ jobs:
       fail-fast: false
       matrix:
         os: [ubuntu-22.04, windows-2022]
-        cuda: ['12.4.0']
+        cuda: ['12.4.0', '13.0.0']
     runs-on: ${{ matrix.os }}
     steps:
       - name: Clone
@@ -417,7 +417,7 @@ jobs:
           ref: '${{ github.event.inputs.llama_cpp_commit }}'
 
       - uses: Jimver/cuda-toolkit@v0.2.15
-        if: runner.os == 'Windows'
+        if: runner.os == 'Windows' && matrix.cuda == '12.4.0'
         id: cuda-toolkit-windows
         with:
           cuda: ${{ matrix.cuda }}
@@ -425,13 +425,30 @@ jobs:
           sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'
 
       - uses: Jimver/cuda-toolkit@v0.2.15
-        if: runner.os == 'Linux'
+        if: runner.os == 'Linux' && matrix.cuda == '12.4.0'
         id: cuda-toolkit-linux
         with:
           cuda: ${{ matrix.cuda }}
           method: 'network'
           linux-local-args: '["--toolkit"]'
 
+      - uses: Jimver/cuda-toolkit@v0.2.36
+        if: runner.os == 'Windows' && matrix.cuda != '12.4.0'
+        id: cuda-toolkit-windows-13
+        with:
+          cuda: ${{ matrix.cuda }}
+          method: 'network'
+          sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'
+          use-github-cache: false
+
+      - uses: Jimver/cuda-toolkit@v0.2.36
+        if: runner.os == 'Linux' && matrix.cuda != '12.4.0'
+        id: cuda-toolkit-linux-13
+        with:
+          cuda: ${{ matrix.cuda }}
+          method: 'network'
+          linux-local-args: '["--toolkit"]'
+
       - name: Build (Windows)
         if: runner.os == 'Windows'
         id: cmake_build_windows
@@ -688,7 +705,7 @@ jobs:
       - name: Rearrange Files
         run: |
           # Make all directories at once
-          mkdir --parents deps/{noavx,avx,avx2,avx512,linux-arm64,musl-noavx,musl-avx,musl-avx2,musl-avx512,osx-arm64,osx-x64,osx-x64-rosetta2,cu12.4.0,vulkan,android-arm64-v8a,android-x86,android-x86_64,win-arm64}
+          mkdir --parents deps/{noavx,avx,avx2,avx512,linux-arm64,musl-noavx,musl-avx,musl-avx2,musl-avx512,osx-arm64,osx-x64,osx-x64-rosetta2,cu12.4.0,cu13.0.0,vulkan,android-arm64-v8a,android-x86,android-x86_64,win-arm64}
 
           # Linux
           cp artifacts/ggml-bin-linux-noavx-x64.so/libggml.so           deps/noavx/libggml.so
@@ -821,6 +838,12 @@ jobs:
           cp artifacts/llama-bin-win-cublas-cu12.4.0-x64.dll/llama.dll          deps/cu12.4.0/llama.dll
           cp artifacts/mtmd-bin-win-cublas-cu12.4.0-x64.dll/mtmd.dll            deps/cu12.4.0/mtmd.dll
 
+          cp artifacts/ggml-bin-win-cublas-cu13.0.0-x64.dll/ggml.dll            deps/cu13.0.0/ggml.dll
+          cp artifacts/ggml-base-bin-win-cublas-cu13.0.0-x64.dll/ggml-base.dll  deps/cu13.0.0/ggml-base.dll
+          cp artifacts/ggml-cuda-bin-win-cublas-cu13.0.0-x64.dll/ggml-cuda.dll  deps/cu13.0.0/ggml-cuda.dll
+          cp artifacts/llama-bin-win-cublas-cu13.0.0-x64.dll/llama.dll          deps/cu13.0.0/llama.dll
+          cp artifacts/mtmd-bin-win-cublas-cu13.0.0-x64.dll/mtmd.dll            deps/cu13.0.0/mtmd.dll
+
           # Linux CUDA
           cp artifacts/ggml-bin-linux-cublas-cu12.4.0-x64.so/libggml.so              deps/cu12.4.0/libggml.so
           cp artifacts/ggml-base-bin-linux-cublas-cu12.4.0-x64.so/libggml-base.so    deps/cu12.4.0/libggml-base.so
@@ -828,6 +851,12 @@ jobs:
           cp artifacts/llama-bin-linux-cublas-cu12.4.0-x64.so/libllama.so            deps/cu12.4.0/libllama.so
           cp artifacts/mtmd-bin-linux-cublas-cu12.4.0-x64.so/libmtmd.so              deps/cu12.4.0/libmtmd.so
 
+          cp artifacts/ggml-bin-linux-cublas-cu13.0.0-x64.so/libggml.so              deps/cu13.0.0/libggml.so
+          cp artifacts/ggml-base-bin-linux-cublas-cu13.0.0-x64.so/libggml-base.so    deps/cu13.0.0/libggml-base.so
+          cp artifacts/ggml-cuda-bin-linux-cublas-cu13.0.0-x64.so/libggml-cuda.so    deps/cu13.0.0/libggml-cuda.so
+          cp artifacts/llama-bin-linux-cublas-cu13.0.0-x64.so/libllama.so            deps/cu13.0.0/libllama.so
+          cp artifacts/mtmd-bin-linux-cublas-cu13.0.0-x64.so/libmtmd.so              deps/cu13.0.0/libmtmd.so
+
           # Windows Vulkan
           cp artifacts/ggml-bin-win-vulkan-x64.dll/ggml.dll               deps/vulkan/ggml.dll
           cp artifacts/ggml-base-bin-win-vulkan-x64.dll/ggml-base.dll     deps/vulkan/ggml-base.dll
diff --git a/LLama.Unittest/NativeLibraryWithCudaTests.cs b/LLama.Unittest/NativeLibraryWithCudaTests.cs
new file mode 100644
index 000000000..a1ae21643
--- /dev/null
+++ b/LLama.Unittest/NativeLibraryWithCudaTests.cs
@@ -0,0 +1,61 @@
+using LLama.Native;
+using System.Runtime.InteropServices;
+
+namespace LLama.Unittest
+{
+    public class NativeLibraryWithCudaTests
+    {
+        [Fact]
+        public void SkipCheckUnknownMajorOnLinuxYieldsCuda13Then12Then11()
+        {
+            var library = new NativeLibraryWithCuda(-1, NativeLibraryName.LLama, AvxLevel.Avx2, true);
+            var systemInfo = new SystemInfo(OSPlatform.Linux, -1, null);
+
+            var paths = library.Prepare(systemInfo, null).ToArray();
+
+            Assert.Equal(3, paths.Length);
+            Assert.Contains("cuda13", paths[0]);
+            Assert.Contains("cuda12", paths[1]);
+            Assert.Contains("cuda11", paths[2]);
+        }
+
+        [Fact]
+        public void DetectedMajor13OnWindowsYieldsCuda13Only()
+        {
+            var library = new NativeLibraryWithCuda(13, NativeLibraryName.LLama, AvxLevel.Avx2, false);
+            var systemInfo = new SystemInfo(OSPlatform.Windows, 13, null);
+
+            var paths = library.Prepare(systemInfo, null).ToArray();
+
+            Assert.Single(paths);
+            Assert.Contains("cuda13", paths[0]);
+            Assert.DoesNotContain("cuda12", paths[0]);
+            Assert.DoesNotContain("cuda11", paths[0]);
+        }
+
+        [Fact]
+        public void DetectedMajor12YieldsCuda12Only()
+        {
+            var library = new NativeLibraryWithCuda(12, NativeLibraryName.LLama, AvxLevel.Avx2, false);
+            var systemInfo = new SystemInfo(OSPlatform.Linux, 12, null);
+
+            var paths = library.Prepare(systemInfo, null).ToArray();
+
+            Assert.Single(paths);
+            Assert.Contains("cuda12", paths[0]);
+            Assert.DoesNotContain("cuda13", paths[0]);
+            Assert.DoesNotContain("cuda11", paths[0]);
+        }
+
+        [Fact]
+        public void SkipCheckFalseUnknownMajorYieldsEmpty()
+        {
+            var library = new NativeLibraryWithCuda(-1, NativeLibraryName.LLama, AvxLevel.Avx2, false);
+            var systemInfo = new SystemInfo(OSPlatform.Linux, -1, null);
+
+            var paths = library.Prepare(systemInfo, null).ToArray();
+
+            Assert.Empty(paths);
+        }
+    }
+}
diff --git a/LLama/Native/Load/NativeLibraryWithCuda.cs b/LLama/Native/Load/NativeLibraryWithCuda.cs
index 36dc4ca81..569d2f930 100644
--- a/LLama/Native/Load/NativeLibraryWithCuda.cs
+++ b/LLama/Native/Load/NativeLibraryWithCuda.cs
@@ -46,7 +46,12 @@ public IEnumerable<string> Prepare(SystemInfo systemInfo, NativeLogConfig.LLamaL
             {
                 if (_majorCudaVersion == -1 && _skipCheck)
                 {
-                    // Currently only 11 and 12 are supported.
+                    // Currently 11, 12, and 13 are supported.
+                    var cuda13LibraryPath = GetCudaPath(systemInfo, 13, logCallback);
+                    if (cuda13LibraryPath is not null)
+                    {
+                        yield return cuda13LibraryPath;
+                    }
                     var cuda12LibraryPath = GetCudaPath(systemInfo, 12, logCallback);
                     if (cuda12LibraryPath is not null)
                     {