diff --git a/python/sglang/srt/models/qwen3_moe.py b/python/sglang/srt/models/qwen3_moe.py
index 1373099fdf5a..010a73074759 100644
--- a/python/sglang/srt/models/qwen3_moe.py
+++ b/python/sglang/srt/models/qwen3_moe.py
@@ -1182,11 +1182,15 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                         logger.warning(f"Parameter {name} not found in params_dict")
 
         if not hasattr(self, "routed_experts_weights_of_layer"):
-            self.routed_experts_weights_of_layer = {
-                layer_id: self.model.layers[layer_id].mlp.get_moe_weights()
-                for layer_id in range(self.start_layer, self.end_layer)
-                if isinstance(self.model.layers[layer_id].mlp, Qwen3MoeSparseMoeBlock)
-            }
+            self.routed_experts_weights_of_layer = LazyValue(
+                lambda: {
+                    layer_id: self.model.layers[layer_id].mlp.get_moe_weights()
+                    for layer_id in range(self.start_layer, self.end_layer)
+                    if isinstance(
+                        self.model.layers[layer_id].mlp, Qwen3MoeSparseMoeBlock
+                    )
+                }
+            )
 
     @classmethod
     def get_model_config_for_expert_location(cls, config):