diff --git a/CHANGELOG.md b/CHANGELOG.md
index 0b03f5e4..f598cc41 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -7,6 +7,14 @@
 
 ## [Unreleased]
 
+### Reliability
+
+- **Cached models no longer retry HuggingFace when offline.** Loading a fully-downloaded
+  model now forces offline mode for the duration of the load, so it skips the network HEAD
+  request (and its 5-retry backoff) for every config file — `config.json`,
+  `generation_config.json`, and the rest — instead of retrying each one in sequence before the
+  app becomes ready.
+
 ### Linux
 
 - **ROCm setup works on Linux AMD systems.** Docker ROCm builds now keep PyTorch
diff --git a/backend/backends/base.py b/backend/backends/base.py
index 70ec11ef..672022e8 100644
--- a/backend/backends/base.py
+++ b/backend/backends/base.py
@@ -14,6 +14,7 @@ from typing import Callable, List, Optional, Tuple
 import numpy as np
 
 from ..utils.audio import normalize_audio, load_audio
+from ..utils.hf_offline_patch import force_offline_if_cached
 from ..utils.progress import get_progress_manager
 from ..utils.hf_progress import HFProgressTracker, create_hf_progress_callback
 from ..utils.tasks import get_task_manager
@@ -280,7 +281,8 @@ def model_load_progress(
         )
 
     try:
-        yield tracker_context
+        with force_offline_if_cached(is_cached, model_name):
+            yield tracker_context
     except Exception as e:
         # Report error to both managers
         progress_manager.mark_error(model_name, str(e))
diff --git a/backend/tests/test_model_load_progress_offline.py b/backend/tests/test_model_load_progress_offline.py
new file mode 100644
index 00000000..4befe58b
--- /dev/null
+++ b/backend/tests/test_model_load_progress_offline.py
@@ -0,0 +1,55 @@
+"""
+Regression test for voicebox #434 (infinite HF retry loop on a cached model).
+
+``model_load_progress`` is the single context manager every backend's
+``_load_model_sync`` uses around its ``from_pretrained()`` call. It already
+receives ``is_cached`` but never forwarded it to ``force_offline_if_cached``,
+so a fully-cached model still resolved config files against huggingface.co
+and ate the default 5-retry backoff per file offline. This asserts the guard
+is actually active for the duration of the ``with`` block when ``is_cached``
+is ``True``, and restored on exit.
+
+NOTE: mutates process-global state in ``huggingface_hub.constants`` and
+``transformers.utils.hub`` (via ``force_offline_if_cached``); run serially,
+same caveat as ``test_offline_guard.py``.
+"""
+
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent.parent.parent))
+
+from backend.backends.base import model_load_progress
+
+
+def _hf_const():
+    import huggingface_hub.constants as hf_const
+
+    return hf_const
+
+
+def _tf_hub():
+    import transformers.utils.hub as tf_hub
+
+    return tf_hub
+
+
+def test_cached_model_load_forces_offline_mode():
+    original_hf = _hf_const().HF_HUB_OFFLINE
+    original_tf = _tf_hub()._is_offline_mode
+
+    with model_load_progress("test-model", is_cached=True):
+        assert _hf_const().HF_HUB_OFFLINE is True
+        assert _tf_hub()._is_offline_mode is True
+
+    assert original_hf == _hf_const().HF_HUB_OFFLINE
+    assert original_tf == _tf_hub()._is_offline_mode
+
+
+def test_uncached_model_load_does_not_force_offline_mode():
+    original_hf = _hf_const().HF_HUB_OFFLINE
+
+    with model_load_progress("test-model", is_cached=False):
+        assert original_hf == _hf_const().HF_HUB_OFFLINE
+
+    assert original_hf == _hf_const().HF_HUB_OFFLINE