diff --git a/slime/backends/megatron_utils/actor.py b/slime/backends/megatron_utils/actor.py
index 586f88ded..0d9f65ec9 100644
--- a/slime/backends/megatron_utils/actor.py
+++ b/slime/backends/megatron_utils/actor.py
@@ -110,6 +110,11 @@ def init(
         # Load teacher model for Megatron-based on-policy distillation
         if with_opd_teacher:
             self.load_other_checkpoint("teacher", args.opd_teacher_load)
+            # Restore actor weights so update_weights() pushes the student (not teacher) to sglang.
+            # Without offload_train the generic recovery below is skipped, leaving self.model
+            # in teacher state and causing step-0 rollouts/evals to run with the teacher model.
+            if not self.args.offload_train:
+                self._switch_model("actor")
 
         if self.args.keep_old_actor:
             # Load old_actor checkpoint