diff --git a/openpilot/selfdrive/modeld/modeld.py b/openpilot/selfdrive/modeld/modeld.py
index 3c7dd935a2..68f49b4113 100755
--- a/openpilot/selfdrive/modeld/modeld.py
+++ b/openpilot/selfdrive/modeld/modeld.py
@@ -359,10 +359,11 @@ def main(demo=False, remote_addr: str | None = None, big_model: bool = False):
 
       fill_driving_model_data(drivingdata_send, modelv2_send)
       fill_pose_msg(posenet_send, model_output, meta_main.frame_id, vipc_dropped_frames, meta_main.timestamp_eof, live_calib_seen)
-      pm.send('modelV2', modelv2_send)
-      pm.send('drivingModelData', drivingdata_send)
-      pm.send('cameraOdometry', posenet_send)
-      pm.send('modelDataV2SP', mdv2sp_send)
+      if remote_addr is not None or not params.get_bool("WgpuEnabled"):
+        pm.send('modelV2', modelv2_send)
+        pm.send('drivingModelData', drivingdata_send)
+        pm.send('cameraOdometry', posenet_send)
+        pm.send('modelDataV2SP', mdv2sp_send)
     last_vipc_frame_id = meta_main.frame_id
 
 
diff --git a/openpilot/system/manager/process_config.py b/openpilot/system/manager/process_config.py
index 9210b99bb8..a72675385d 100644
--- a/openpilot/system/manager/process_config.py
+++ b/openpilot/system/manager/process_config.py
@@ -95,9 +95,6 @@ def is_stock_model(started, params, CP: car.CarParams) -> bool:
   """Check if the active model runner is stock."""
   return bool(get_active_model_runner(params, not started) == custom.ModelManagerSP.Runner.stock)
 
-def not_wgpu(started: bool, params: Params, CP: car.CarParams) -> bool:
-  return not params.get_bool("WgpuEnabled")
-
 def mapd_ready(started: bool, params: Params, CP: car.CarParams) -> bool:
   return bool(os.path.exists(Paths.mapd_root()))
 
@@ -131,7 +128,7 @@ procs = [
   PythonProcess("micd", "openpilot.system.micd", iscar),
   PythonProcess("timed", "openpilot.system.timed", always_run, enabled=not PC),
 
-  PythonProcess("modeld", "openpilot.selfdrive.modeld.modeld", and_(and_(only_onroad, is_stock_model), not_wgpu)),
+  PythonProcess("modeld", "openpilot.selfdrive.modeld.modeld", and_(only_onroad, is_stock_model)),
   PythonProcess("dmonitoringmodeld", "openpilot.selfdrive.modeld.dmonitoringmodeld", driverview, enabled=(WEBCAM or not PC)),
 
   PythonProcess("sensord", "openpilot.system.sensord.sensord", only_onroad, enabled=not PC),
diff --git a/openpilot/tools/wgpu/README.md b/openpilot/tools/wgpu/README.md
index 89892913d0..25208b026b 100644
--- a/openpilot/tools/wgpu/README.md
+++ b/openpilot/tools/wgpu/README.md
@@ -4,9 +4,10 @@ This runs driving `modeld` on a laptop and returns its cereal outputs to a comma
 device over the existing Wi-Fi network. It reuses the existing HEVC camera
 stream, VisionIPC decoder, and cereal ZMQ bridge.
 
-This is for offroad/bench testing only. Wi-Fi has no deterministic latency or
-availability guarantee. The device restores local `modeld` when the device-side
-helper exits, but that is not a seamless onroad failover.
+This is for controlled bench testing only. Wi-Fi has no deterministic latency
+or availability guarantee. Local `modeld` remains warm during a WGPU session;
+the device-side helper switches only after receiving a fresh remote model and
+restores local publication if the remote model is missing for 350 ms.
 
 ## Build
 
@@ -38,8 +39,9 @@ the small model on that class of laptop.
 
 ## Run
 
-Find the laptop's LAN IP address that the comma device can reach. While the
-device is offroad, run:
+Find the laptop's LAN IP address that the comma device can reach. The helper can
+start while onroad: it forwards camera/state while local `modeld` remains active,
+then switches publication after the laptop produces a fresh valid model:
 
 ```sh
 cd /data/openpilot
@@ -56,7 +58,9 @@ python3 -m openpilot.tools.wgpu.host COMMA_IP
 Add `--big-model` after `COMMA_IP` to use the locally compiled big model.
 
 The first remote `carParams` packet can take up to 50 seconds. Stop either side
-with Ctrl+C. Stop the device helper before changing branches or rebooting.
+with Ctrl+C. A host disconnect automatically restores the warm local publisher
+after a 350 ms timeout. Stop the device helper before changing branches or
+rebooting.
 
 For stationary bench testing only, model lag can be changed from a blocking
 event to a live warning showing model age and dropped frames. After ignition is
diff --git a/openpilot/tools/wgpu/device.py b/openpilot/tools/wgpu/device.py
index 7be206f2e0..1d7e5820f2 100644
--- a/openpilot/tools/wgpu/device.py
+++ b/openpilot/tools/wgpu/device.py
@@ -5,10 +5,14 @@ import subprocess
 import time
 from pathlib import Path
 
+import openpilot.cereal.messaging as messaging
 from openpilot.common.params import Params
+from openpilot.common.realtime import DT_MDL
+from openpilot.tools.wgpu.zmq import ZmqSubSocket
 
 
 MODEL_OUTPUTS = "modelV2,drivingModelData,cameraOdometry,modelDataV2SP"
+REMOTE_MODEL_TIMEOUT = 0.35
 ROOT = Path(__file__).resolve().parents[3]
 BRIDGE = ROOT / "openpilot/cereal/messaging/bridge"
 
@@ -26,6 +30,17 @@ def handle_sigterm(*_) -> None:
   raise KeyboardInterrupt
 
 
+def receive_fresh_model(sock: ZmqSubSocket) -> bool:
+  raw = sock.receive(non_blocking=True)
+  if raw is None:
+    return False
+  event = messaging.log_from_bytes(raw)
+  if event.which() != "modelV2" or not event.valid:
+    return False
+  model_age = (time.monotonic_ns() - event.modelV2.timestampEof) / 1e9
+  return 0 <= model_age < REMOTE_MODEL_TIMEOUT
+
+
 def main() -> None:
   parser = argparse.ArgumentParser(description="Route modeld traffic between this device and a wireless host.")
   parser.add_argument("host", help="Laptop IP address reachable from this device")
@@ -35,27 +50,45 @@ def main() -> None:
     raise FileNotFoundError(f"build the cereal bridge first: {BRIDGE}")
 
   params = Params()
-  if not params.get_bool("IsOffroad"):
-    raise RuntimeError("start the wgpu bridge while offroad")
-
-  procs: list[subprocess.Popen] = []
+  forward: subprocess.Popen | None = None
+  reverse: subprocess.Popen | None = None
+  wgpu_enabled = False
   try:
+    # Keep local modeld publishing while the remote model connects and warms up.
+    forward = subprocess.Popen([str(BRIDGE)])
+    remote_model = ZmqSubSocket("modelV2", args.host, conflate=True)
+    print(f"forwarding camera/state to {args.host}; waiting for a fresh remote model")
+    while not receive_fresh_model(remote_model):
+      if forward.poll() is not None:
+        raise RuntimeError(f"forward bridge exited with status {forward.returncode}")
+      time.sleep(0.05)
+
+    # Local modeld remains warm but stops publishing when this flag changes.
     params.put_bool("WgpuEnabled", True, block=True)
-    procs = [
-      subprocess.Popen([str(BRIDGE)]),
-      subprocess.Popen([str(BRIDGE), args.host, MODEL_OUTPUTS]),
-    ]
-    print(f"wgpu enabled; forwarding camera/state to {args.host}")
-    print("keep this process running; Ctrl+C restores local modeld")
-    while all(proc.poll() is None for proc in procs):
-      time.sleep(0.25)
-    failed = next(proc for proc in procs if proc.poll() is not None)
-    raise RuntimeError(f"bridge exited with status {failed.returncode}")
+    wgpu_enabled = True
+    time.sleep(2 * DT_MDL)
+    reverse = subprocess.Popen([str(BRIDGE), args.host, MODEL_OUTPUTS])
+    print("wgpu active; Ctrl+C or loss of the remote model restores local modeld")
+
+    last_remote_model = time.monotonic()
+    while True:
+      if receive_fresh_model(remote_model):
+        last_remote_model = time.monotonic()
+      if forward.poll() is not None:
+        raise RuntimeError(f"forward bridge exited with status {forward.returncode}")
+      if reverse.poll() is not None:
+        raise RuntimeError(f"reverse bridge exited with status {reverse.returncode}")
+      if time.monotonic() - last_remote_model > REMOTE_MODEL_TIMEOUT:
+        raise RuntimeError("remote model timed out; restoring local modeld")
+      time.sleep(0.05)
   finally:
-    params.put_bool("WgpuEnabled", False, block=True)
-    for proc in procs:
-      if proc.poll() is None:
-        stop_process(proc)
+    # Stop remote publication before allowing the warm local publisher to resume.
+    if reverse is not None and reverse.poll() is None:
+      stop_process(reverse)
+    if wgpu_enabled:
+      params.put_bool("WgpuEnabled", False, block=True)
+    if forward is not None and forward.poll() is None:
+      stop_process(forward)
     print("wgpu disabled; local modeld restored")