diff --git a/vllm/envs.py b/vllm/envs.py
index 689428ec59..0833949b52 100755
--- a/vllm/envs.py
+++ b/vllm/envs.py
@@ -64,6 +64,7 @@ if TYPE_CHECKING:
     VLLM_XLA_USE_SPMD: bool = False
     VLLM_WORKER_MULTIPROC_METHOD: Literal["fork", "spawn"] = "fork"
     VLLM_ASSETS_CACHE: str = os.path.join(VLLM_CACHE_ROOT, "assets")
+    VLLM_ASSETS_CACHE_MODEL_CLEAN: bool = False
     VLLM_IMAGE_FETCH_TIMEOUT: int = 5
     VLLM_VIDEO_FETCH_TIMEOUT: int = 30
     VLLM_AUDIO_FETCH_TIMEOUT: int = 10
@@ -699,6 +700,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
             os.path.join(get_default_cache_root(), "vllm", "assets"),
         )),
 
+    # If the env var is set, we will clean model file in
+    # this path $VLLM_ASSETS_CACHE/model_streamer/$model_name
+    "VLLM_ASSETS_CACHE_MODEL_CLEAN":
+    lambda: bool(int(os.getenv("VLLM_ASSETS_CACHE_MODEL_CLEAN", "0"))),
+
     # Timeout for fetching images when serving multimodal models
     # Default is 5 seconds
     "VLLM_IMAGE_FETCH_TIMEOUT":
diff --git a/vllm/transformers_utils/runai_utils.py b/vllm/transformers_utils/runai_utils.py
index 08466ca19b..355fd60e8d 100644
--- a/vllm/transformers_utils/runai_utils.py
+++ b/vllm/transformers_utils/runai_utils.py
@@ -5,9 +5,10 @@ import hashlib
 import os
 import shutil
 import signal
-import tempfile
 from typing import Optional
 
+from vllm import envs
+from vllm.assets.base import get_cache_dir
 from vllm.logger import init_logger
 from vllm.utils import PlaceholderModule
 
@@ -58,20 +59,19 @@ class ObjectStorageModel:
     """
 
     def __init__(self, url: str) -> None:
-        for sig in (signal.SIGINT, signal.SIGTERM):
-            existing_handler = signal.getsignal(sig)
-            signal.signal(sig, self._close_by_signal(existing_handler))
+        if envs.VLLM_ASSETS_CACHE_MODEL_CLEAN:
+            for sig in (signal.SIGINT, signal.SIGTERM):
+                existing_handler = signal.getsignal(sig)
+                signal.signal(sig, self._close_by_signal(existing_handler))
 
         dir_name = os.path.join(
-            tempfile.gettempdir(),
+            get_cache_dir(), "model_streamer",
             hashlib.sha256(str(url).encode()).hexdigest()[:8])
         if os.path.exists(dir_name):
             shutil.rmtree(dir_name)
         os.makedirs(dir_name)
         self.dir = dir_name
-
-    def __del__(self):
-        self._close()
+        logger.debug("Init object storage, model cache path is: %s", dir_name)
 
     def _close(self) -> None:
         if os.path.exists(self.dir):