diff --git a/.github/workflows/wheel.yml b/.github/workflows/wheel.yml
index a6c7f2f..0fc0f4d 100644
--- a/.github/workflows/wheel.yml
+++ b/.github/workflows/wheel.yml
@@ -58,6 +58,33 @@ jobs:
           name: wheel-${{ matrix.wheel_tag }}
           path: wheelhouse/pygpubench*.whl
 
+  gpu-test:
+    name: GPU tests (Modal L4)
+    needs: wheel
+    runs-on: ubuntu-24.04
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Download wheel
+        uses: actions/download-artifact@v4
+        with:
+          pattern: wheel-*
+          path: dist/
+          merge-multiple: true
+
+      - name: Install Modal
+        run: pip install modal
+
+      - name: Run GPU tests
+        env:
+          MODAL_TOKEN_ID: ${{ secrets.MODAL_TOKEN_ID }}
+          MODAL_TOKEN_SECRET: ${{ secrets.MODAL_TOKEN_SECRET }}
+        run: modal run ci/modal_gpu_test.py --wheel dist/pygpubench*.whl --test-dir test
+
   release:
     name: Publish to GitHub Releases
     needs: wheel
diff --git a/ci/modal_gpu_test.py b/ci/modal_gpu_test.py
new file mode 100644
index 0000000..0e30399
--- /dev/null
+++ b/ci/modal_gpu_test.py
@@ -0,0 +1,69 @@
+"""Run pygpubench GPU tests on a Modal L4 GPU.
+
+Usage: modal run ci/modal_gpu_test.py <wheel_path> <test_dir>
+"""
+
+import modal
+from pathlib import Path
+
+image = (
+    modal.Image.from_registry(
+        "nvidia/cuda:13.0.1-cudnn-devel-ubuntu24.04", add_python="3.12"
+    )
+    .entrypoint([])
+    .uv_pip_install("torch", index_url="https://download.pytorch.org/whl/cu130")
+)
+
+app = modal.App("pygpubench-ci", image=image)
+
+
+@app.function(gpu="L4", timeout=600)
+def run_tests(whl_bytes: bytes, whl_name: str, test_files: dict[str, bytes]):
+    import subprocess
+    import sys
+    import os
+
+    # Write wheel and install it
+    whl_path = f"/tmp/{whl_name}"
+    with open(whl_path, "wb") as f:
+        f.write(whl_bytes)
+    subprocess.run([sys.executable, "-m", "pip", "install", whl_path], check=True)
+
+    # Write test files
+    test_dir = "/tmp/tests"
+    os.makedirs(test_dir, exist_ok=True)
+    for name, content in test_files.items():
+        with open(os.path.join(test_dir, name), "wb") as f:
+            f.write(content)
+
+    # Run all test scripts
+    os.chdir(test_dir)
+    failed = []
+    for test_file in sorted(test_files):
+        if test_file == "submission.py":
+            continue
+        print(f"\n=== {test_file} ===")
+        result = subprocess.run([sys.executable, test_file], text=True)
+        if result.returncode != 0:
+            failed.append(test_file)
+
+    if failed:
+        print(f"\nFailed: {', '.join(failed)}")
+        raise SystemExit(1)
+
+
+@app.local_entrypoint()
+def main(wheel: str, test_dir: str = "test"):
+    import glob
+
+    # Read the wheel
+    whl_path = Path(wheel)
+    whl_bytes = whl_path.read_bytes()
+
+    # Read all test files
+    test_path = Path(test_dir)
+    test_files = {}
+    for f in test_path.glob("*.py"):
+        test_files[f.name] = f.read_bytes()
+
+    run_tests.remote(whl_bytes, whl_path.name, test_files)