diff --git a/selfdrive/selfdrived/selfdrived.py b/selfdrive/selfdrived/selfdrived.py
index 86cc43d1f9..b33cd627ae 100644
--- a/selfdrive/selfdrived/selfdrived.py
+++ b/selfdrive/selfdrived/selfdrived.py
@@ -707,13 +707,21 @@ class SelfdriveD:
                          (contains_event_type(self.events, self.starpilot_events, ET.SOFT_DISABLE) or
                           contains_event_type(self.events, self.starpilot_events, ET.IMMEDIATE_DISABLE))
     no_system_errors = (not has_disable_events) or (len(self.events) == num_events)
-    # modeld publishes on the small model throughout the background load, so only the model
-    # swap itself can briefly disturb the stream. Suppressing for the whole load would hide
-    # a genuinely dead modeld for as long as the load takes.
+    # The model swap briefly disturbs the stream, so forgive everything around it.
     big_model_settling = time.monotonic() < self.big_model_swap_t + 2.
     all_checks = self.sm.all_checks()
     all_alive = self.sm.all_alive() if not all_checks else True
     all_freq_ok = self.sm.all_freq_ok() if not all_checks else True
+    # Loading the big model realizes its graph on the same QCOM GPU the small model drives
+    # on, which costs modeld frames for part of the load and drags modelV2 below its target
+    # rate. commIssueAvgFreq is NO_ENTRY, so that would block engaging on a small model that
+    # is working fine. Forgive only a slow modelV2, and only while a load is in flight --
+    # anything dying, and any other service, still reports normally.
+    if self.big_model_loading and not all_checks and all_alive and not all_freq_ok:
+      slow = {s for s, freq_ok in self.sm.freq_ok.items() if not freq_ok}
+      if slow and slow <= {'modelV2', 'drivingModelData', 'cameraOdometry'} and self.sm.all_valid():
+        all_freq_ok = True
+        all_checks = True
     report_comm_issue, self.valid_only_comm_issue_frames = evaluate_comm_issue(
       all_checks, all_alive, all_freq_ok, self.valid_only_comm_issue_frames,
     )
diff --git a/selfdrive/selfdrived/tests/test_big_model_engagement.py b/selfdrive/selfdrived/tests/test_big_model_engagement.py
index bde9410e93..45f7fc17ea 100644
--- a/selfdrive/selfdrived/tests/test_big_model_engagement.py
+++ b/selfdrive/selfdrived/tests/test_big_model_engagement.py
@@ -68,6 +68,30 @@ def test_big_model_loading_raises_no_alert():
   assert EVENTS[EventName.bigModelLoading] == {}
 
 
+def test_slow_modelv2_during_a_load_does_not_block_engaging():
+  """commIssueAvgFreq is NO_ENTRY, so a load-induced modelV2 slowdown blocks engaging.
+
+  Measured on drive 00000abd: modelV2 p95 hit 732 ms during the load while roadCameraState
+  stayed at 51 ms, and the resulting commIssue/commIssueAvgFreq storm ran 38.5 s - 47.9 s.
+  The forgiveness must be narrow: only a *slow* modelV2 family, only while loading, and only
+  when everything is still alive and valid.
+  """
+  from openpilot.selfdrive.selfdrived import selfdrived
+
+  source = (Path(selfdrived.__file__)).read_text(encoding="utf-8")
+  guard = next(
+    node for node in ast.walk(ast.parse(source))
+    if isinstance(node, ast.If) and "big_model_loading" in ast.dump(node.test)
+    and "all_freq_ok" in ast.dump(node.test)
+  )
+  test = ast.dump(guard.test)
+  # A dead service must still be reported, so aliveness is required, not forgiven.
+  assert "all_alive" in test
+  body = ast.dump(guard)
+  assert "all_valid" in body, "stale/invalid services must still raise commIssue"
+  assert "modelV2" in body, "only the modeld output services may be forgiven"
+
+
 def test_frame_drop_warning_is_suppressed_while_the_big_model_loads():
   """Loading realizes the big model's graph on QCOM, the GPU the small model drives on.