diff --git a/cmake/utils.cmake b/cmake/utils.cmake
index 3be7655117..ed41ab430e 100644
--- a/cmake/utils.cmake
+++ b/cmake/utils.cmake
@@ -421,7 +421,13 @@ function (define_gpu_extension_target GPU_MOD_NAME)
     set(GPU_WITH_SOABI)
   endif()
 
-  if (GPU_USE_SABI)
+  run_python(IS_FREETHREADED_PYTHON
+    "import sysconfig; print(1 if sysconfig.get_config_var(\"Py_GIL_DISABLED\") else 0)"
+    "Failed to determine whether interpreter is free-threaded")
+
+  # Free-threaded Python doesn't yet support the stable ABI (see PEP 803/809),
+  # so avoid using the stable ABI under free-threading only.
+  if (GPU_USE_SABI AND NOT IS_FREETHREADED_PYTHON)
     Python_add_library(${GPU_MOD_NAME} MODULE USE_SABI ${GPU_USE_SABI} ${GPU_WITH_SOABI} "${GPU_SOURCES}")
   else()
     Python_add_library(${GPU_MOD_NAME} MODULE ${GPU_WITH_SOABI} "${GPU_SOURCES}")
diff --git a/setup.py b/setup.py
index 00d482604c..d2479f600f 100644
--- a/setup.py
+++ b/setup.py
@@ -4,6 +4,7 @@
 import os
 import re
 import ast
+import sysconfig
 from collections import namedtuple
 from pathlib import Path
 from typing import Dict
@@ -79,10 +80,14 @@ def _is_hip() -> bool:
             or VLLM_TARGET_DEVICE == "rocm") and torch.version.hip is not None
 
 
+def is_freethreaded():
+    return bool(sysconfig.get_config_var("Py_GIL_DISABLED"))
+
+
 class CMakeExtension(Extension):
 
     def __init__(self, name: str, cmake_lists_dir: str = '.', **kwa) -> None:
-        super().__init__(name, sources=[], py_limited_api=True, **kwa)
+        super().__init__(name, sources=[], py_limited_api=not is_freethreaded(), **kwa)
         self.cmake_lists_dir = os.path.abspath(cmake_lists_dir)