diff --git a/common/params_keys.h b/common/params_keys.h
index dcd46dbd6a..4a9fd1c775 100644
--- a/common/params_keys.h
+++ b/common/params_keys.h
@@ -352,6 +352,7 @@ inline static std::unordered_map<std::string, ParamKeyAttributes> keys = {
     {"DrivingModelVersion", {PERSISTENT, STRING, "v15", "v15", 1}},
     {"DynamicPathWidth", {PERSISTENT, BOOL, "0", "0", 2, SETTINGS_SIMPLE}},
     {"DynamicPedalsOnUI", {PERSISTENT, BOOL, "1", "0", 1, SETTINGS_SIMPLE}},
+    {"GpuModelReadySound", {PERSISTENT, BOOL, "1", "1", 2, SETTINGS_SIMPLE}},
     {"EngageVolume", {PERSISTENT, INT, "101", "101", 2, SETTINGS_SIMPLE}},
     {"EVTuning", {PERSISTENT, BOOL, "0", "0", 3}},
     {"Fahrenheit", {PERSISTENT, BOOL, "0", "0", 3}},
diff --git a/selfdrive/assets/sounds/model_ready.wav b/selfdrive/assets/sounds/model_ready.wav
new file mode 100644
index 0000000000..b0592f6573
Binary files /dev/null and b/selfdrive/assets/sounds/model_ready.wav differ
diff --git a/selfdrive/ui/soundd.py b/selfdrive/ui/soundd.py
index 360c6f77df..e2ede4fdf4 100644
--- a/selfdrive/ui/soundd.py
+++ b/selfdrive/ui/soundd.py
@@ -42,6 +42,8 @@ StarPilotAudibleAlert = custom.StarPilotCarControl.HUDControl.AudibleAlert
 STARPILOT_CUSTOM_ALERT_OFFSET = 1000
 STARPILOT_CUSTOM_ALERT_START = int(StarPilotAudibleAlert.angry)
 TURN_STEERING_LIMIT_ALERT_SUFFIX = "steersaturated"
+GPU_MODEL_READY_ALERT = 2000
+
 # Keep carState out of this list; C4's onroad stack is near msgq's 15-reader limit.
 SOUNDD_SERVICES = ('selfdriveState', 'soundPressure', 'starpilotSelfdriveState', 'starpilotPlan')
 
@@ -67,6 +69,7 @@ def should_mute_turn_steering_limit_alert(alert_type: str, v_ego: float, mute_be
 
 
 sound_list: dict[int, tuple[str, int | None, float]] = {
+  GPU_MODEL_READY_ALERT: ("model_ready.wav", 1, MAX_VOLUME),
   # AudibleAlert, file name, play count (none for infinite)
   AudibleAlert.engage: ("engage.wav", 1, MAX_VOLUME),
   AudibleAlert.disengage: ("disengage.wav", 1, MAX_VOLUME),
@@ -124,6 +127,11 @@ class Soundd:
     self.spl_filter_weighted = FirstOrderFilter(0, 2.5, FILTER_DT, initialized=False)
 
     self.params_memory = Params(memory=True)
+    from openpilot.starpilot.common.gpu_model_ready_sound import GpuModelReadyChime
+    self.model_ready_chime = GpuModelReadyChime()
+    self.model_ready_params = Params(return_defaults=True)
+    self.model_ready_last_check = 0.0
+    self.model_ready_pending = False
 
     self.starpilot_toggles = get_starpilot_toggles()
 
@@ -325,6 +333,33 @@ class Soundd:
       self.update_alert(AudibleAlert.none)
       self.selfdrive_timeout_alert = False
 
+  def update_model_ready_sound(self, sm):
+    # A one-shot ending exactly on a callback boundary must release its slot.
+    if self.current_alert == GPU_MODEL_READY_ALERT:
+      loaded = self.loaded_sounds.get(GPU_MODEL_READY_ALERT)
+      if loaded is None or self.current_sound_frame >= len(loaded):
+        self.current_alert = AudibleAlert.none
+        self.current_sound_frame = 0
+    now = time.monotonic()
+    if now - self.model_ready_last_check >= 0.25:
+      self.model_ready_last_check = now
+      try:
+        self.model_ready_pending = self.model_ready_chime.update(
+          active=self.model_ready_params.get_bool("UsbGpuActive"),
+          loading=self.model_ready_params.get_bool("UsbGpuLoading"),
+          onroad=self.model_ready_params.get_bool("IsOnroad"),
+          enabled=self.model_ready_params.get_bool("GpuModelReadySound"), now=now)
+      except Exception:
+        self.model_ready_chime.consume()
+        self.model_ready_pending = False
+    # Run after stock/custom alert selection: warnings and timeout alerts always win.
+    if (self.model_ready_pending and self.current_alert == AudibleAlert.none
+        and not self.selfdrive_timeout_alert and sm.valid["selfdriveState"] and sm.alive["selfdriveState"]):
+      self.model_ready_chime.consume()
+      self.model_ready_pending = False
+      self.current_alert_type = ""
+      self.update_alert(GPU_MODEL_READY_ALERT)
+
   def get_volume_override(self):
     if self.current_alert_type.startswith("belowSteerSpeed/"):
       return self.starpilot_toggles.below_steer_speed_volume / 100.0
@@ -391,6 +426,7 @@ class Soundd:
               self.current_volume = 0.0
 
           self.get_audible_alert(sm)
+          self.update_model_ready_sound(sm)
 
           if self.current_alert != AudibleAlert.none:
             v_ego = max(float(getattr(sm["starpilotSelfdriveState"], "vEgo", 0.0)), 0.0)
diff --git a/selfdrive/ui/tests/test_gpu_model_ready_sound.py b/selfdrive/ui/tests/test_gpu_model_ready_sound.py
new file mode 100644
index 0000000000..1911ea5ab3
--- /dev/null
+++ b/selfdrive/ui/tests/test_gpu_model_ready_sound.py
@@ -0,0 +1,70 @@
+import ast
+from pathlib import Path
+from types import SimpleNamespace
+import wave
+
+import pytest
+
+from openpilot.starpilot.common.gpu_model_ready_sound import GpuModelReadyChime
+
+
+def update(chime, now=0, **kwargs):
+  return chime.update(**dict(active=False, loading=False, onroad=True, enabled=True, now=now) | kwargs)
+
+
+def test_ready_once_after_success_not_during_loading_or_failure():
+  chime = GpuModelReadyChime()
+  assert not update(chime, loading=True)
+  assert not update(chime, 1, active=True, loading=True)
+  assert update(chime, 2, active=True)
+  chime.consume()
+  assert not update(chime, 3, active=True)
+  assert not update(chime, 4, loading=True)
+  assert not update(chime, 5)
+
+
+def test_restart_with_already_ready_model_does_not_chime():
+  assert not update(GpuModelReadyChime(), active=True)
+
+
+@pytest.mark.parametrize("cancel", [dict(enabled=False), dict(onroad=False), dict(active=False), dict(loading=True)])
+def test_cancel_pending_on_disabled_offroad_fallback_or_reload(cancel):
+  chime = GpuModelReadyChime()
+  update(chime)
+  assert update(chime, 1, active=True)
+  assert not update(chime, 2, **(dict(active=True) | cancel))
+
+
+def test_pending_notification_expires_without_late_chime():
+  chime = GpuModelReadyChime()
+  update(chime)
+  assert update(chime, 1, active=True)
+  assert not update(chime, 6, active=True)
+  assert not update(chime, 7, active=True)
+
+
+def test_enabling_option_after_load_does_not_replay():
+  chime = GpuModelReadyChime()
+  update(chime, enabled=False)
+  assert not update(chime, 1, active=True, enabled=False)
+  assert not update(chime, 2, active=True, enabled=True)
+
+
+@pytest.mark.parametrize("alert,timeout,valid,alive,expected", [(1,False,True,True,False),(0,True,True,True,False),(0,False,False,True,False),(0,False,True,False,False),(0,False,True,True,True)])
+def test_soundd_never_replaces_existing_or_timeout_alerts(alert,timeout,valid,alive,expected):
+  source=(Path(__file__).parents[1]/'soundd.py').read_text()
+  cls=next(n for n in ast.parse(source).body if isinstance(n,ast.ClassDef) and n.name=='Soundd')
+  method=next(n for n in cls.body if isinstance(n,ast.FunctionDef) and n.name=='update_model_ready_sound')
+  scope={'time':SimpleNamespace(monotonic=lambda:1.1),'AudibleAlert':SimpleNamespace(none=0),'GPU_MODEL_READY_ALERT':2000}
+  exec(compile(ast.Module(body=[method],type_ignores=[]),'<soundd method>','exec'),scope)
+  calls=[]
+  sound=SimpleNamespace(model_ready_last_check=1.0,model_ready_pending=True,model_ready_chime=GpuModelReadyChime(),current_alert=alert,selfdrive_timeout_alert=timeout,update_alert=calls.append)
+  sm=SimpleNamespace(valid={'selfdriveState':valid},alive={'selfdriveState':alive})
+  scope['update_model_ready_sound'](sound,sm)
+  assert calls == ([2000] if expected else [])
+
+
+def test_notification_asset_matches_soundd_audio_format():
+  with wave.open(str(Path(__file__).parents[2]/'assets/sounds/model_ready.wav')) as audio:
+    assert (audio.getnchannels(),audio.getsampwidth(),audio.getframerate()) == (1,2,48000)
+    assert 0 < audio.getnframes() < 48000
diff --git a/starpilot/common/assets/device_settings_layout.json b/starpilot/common/assets/device_settings_layout.json
index 27ddec6a37..a1fba4454c 100644
--- a/starpilot/common/assets/device_settings_layout.json
+++ b/starpilot/common/assets/device_settings_layout.json
@@ -3355,6 +3355,14 @@
         "data_type": "bool",
         "ui_type": "toggle",
         "settings_tier": "simple"
+      },
+      {
+        "key": "GpuModelReadySound",
+        "label": "GPU Model Ready Sound",
+        "description": "Play a short chime when the external GPU model finishes loading. Driving alerts take priority.",
+        "data_type": "bool",
+        "ui_type": "toggle",
+        "settings_tier": "simple"
       }
     ]
   },
diff --git a/starpilot/common/gpu_model_ready_sound.py b/starpilot/common/gpu_model_ready_sound.py
new file mode 100644
index 0000000000..3e4d8803d4
--- /dev/null
+++ b/starpilot/common/gpu_model_ready_sound.py
@@ -0,0 +1,20 @@
+"""Optional notification state; never supersedes driving alerts or starts inference."""
+
+
+class GpuModelReadyChime:
+  def __init__(self):
+    self.ready = None
+    self.pending_until = None
+
+  def update(self, *, active, loading, onroad, enabled, now):
+    ready = bool(onroad and active and not loading)
+    # A soundd restart with an already-loaded model is not a loading event.
+    if self.ready is False and ready and enabled:
+      self.pending_until = now + 5.0
+    self.ready = ready
+    if not enabled or not ready or (self.pending_until is not None and now >= self.pending_until):
+      self.pending_until = None
+    return self.pending_until is not None
+
+  def consume(self):
+    self.pending_until = None