diff --git a/vllm/model_executor/models/qwen3_5.py b/vllm/model_executor/models/qwen3_5.py
index 03ea538127a..2449724cd04 100644
--- a/vllm/model_executor/models/qwen3_5.py
+++ b/vllm/model_executor/models/qwen3_5.py
@@ -306,9 +306,12 @@ class Qwen3_5Model(Qwen3NextModel):
         loaded_params: set[str] = set()
         expert_params_mapping = self.get_expert_mapping()
         is_fused_expert = False
+        base_layer = (
+            "base_layer." if any(".base_layer." in name for name in params_dict) else ""
+        )
         fused_expert_params_mapping = [
-            ("experts.w13_weight", "experts.gate_up_proj", 0, "w1"),
-            ("experts.w2_weight", "experts.down_proj", 0, "w2"),
+            (f"experts.{base_layer}w13_weight", "experts.gate_up_proj", 0, "w1"),
+            (f"experts.{base_layer}w2_weight", "experts.down_proj", 0, "w2"),
         ]
         num_experts = (
             self.config.num_experts if hasattr(self.config, "num_experts") else 0
diff --git a/vllm/model_executor/models/qwen3_5_mtp.py b/vllm/model_executor/models/qwen3_5_mtp.py
index e49806365e3..bbb296d28c9 100644
--- a/vllm/model_executor/models/qwen3_5_mtp.py
+++ b/vllm/model_executor/models/qwen3_5_mtp.py
@@ -207,9 +207,12 @@ class Qwen3_5MultiTokenPredictor(nn.Module):
         params_dict = dict(self.named_parameters())
         loaded_params: set[str] = set()
         is_fused_expert = False
+        base_layer = (
+            "base_layer." if any(".base_layer." in name for name in params_dict) else ""
+        )
         fused_expert_params_mapping = [
-            ("experts.w13_weight", "experts.gate_up_proj", 0, "w1"),
-            ("experts.w2_weight", "experts.down_proj", 0, "w2"),
+            (f"experts.{base_layer}w13_weight", "experts.gate_up_proj", 0, "w1"),
+            (f"experts.{base_layer}w2_weight", "experts.down_proj", 0, "w2"),
         ]
         num_experts = (
             self.config.num_experts if hasattr(self.config, "num_experts") else 0
diff --git a/vllm/model_executor/models/qwen3_vl_moe.py b/vllm/model_executor/models/qwen3_vl_moe.py
index a9c01ccf595..7a3cda1ccbf 100644
--- a/vllm/model_executor/models/qwen3_vl_moe.py
+++ b/vllm/model_executor/models/qwen3_vl_moe.py
@@ -183,9 +183,12 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):
         loaded_params: set[str] = set()
         expert_params_mapping = self.get_expert_mapping()
         is_fused_expert = False
+        base_layer = (
+            "base_layer." if any(".base_layer." in name for name in params_dict) else ""
+        )
         fused_expert_params_mapping = [
-            ("experts.w13_weight", "experts.gate_up_proj", 0, "w1"),
-            ("experts.w2_weight", "experts.down_proj", 0, "w2"),
+            (f"experts.{base_layer}w13_weight", "experts.gate_up_proj", 0, "w1"),
+            (f"experts.{base_layer}w2_weight", "experts.down_proj", 0, "w2"),
         ]
         num_experts = self.config.num_experts
         for name, loaded_weight in weights:
diff --git a/vllm/model_executor/models/step3_text.py b/vllm/model_executor/models/step3_text.py
index 636a121c590..912a1b07546 100644
--- a/vllm/model_executor/models/step3_text.py
+++ b/vllm/model_executor/models/step3_text.py
@@ -463,11 +463,14 @@ class Step3TextForCausalLM(nn.Module, SupportsPP):
         ]
         params_dict = dict(self.named_parameters())
         loaded_params: set[str] = set()
+        base_layer = (
+            "base_layer." if any(".base_layer." in name for name in params_dict) else ""
+        )
 
         expert_params_mapping = [
-            (".moe.experts.w13_weight", ".moe.gate_proj.weight", "w1"),
-            (".moe.experts.w13_weight", ".moe.up_proj.weight", "w3"),
-            (".moe.experts.w2_weight", ".moe.down_proj.weight", "w2"),
+            (f".moe.experts.{base_layer}w13_weight", ".moe.gate_proj.weight", "w1"),
+            (f".moe.experts.{base_layer}w13_weight", ".moe.up_proj.weight", "w3"),
+            (f".moe.experts.{base_layer}w2_weight", ".moe.down_proj.weight", "w2"),
         ]
 
         disable_moe_stacked_params = [data[1] for data in expert_params_mapping]
diff --git a/vllm/model_executor/models/step3p5.py b/vllm/model_executor/models/step3p5.py
index 018f7895602..8b53c657b1e 100644
--- a/vllm/model_executor/models/step3p5.py
+++ b/vllm/model_executor/models/step3p5.py
@@ -626,12 +626,15 @@ class Step3p5Model(nn.Module):
 
         params_dict = dict(self.named_parameters())
         loaded_params: set[str] = set()
+        base_layer = (
+            "base_layer." if any(".base_layer." in name for name in params_dict) else ""
+        )
 
         # Old packed 3D format: .moe.gate_proj.weight [num_experts, out, in]
         expert_params_mapping = [
-            (".moe.experts.w13_weight", ".moe.gate_proj.weight", "w1"),
-            (".moe.experts.w13_weight", ".moe.up_proj.weight", "w3"),
-            (".moe.experts.w2_weight", ".moe.down_proj.weight", "w2"),
+            (f".moe.experts.{base_layer}w13_weight", ".moe.gate_proj.weight", "w1"),
+            (f".moe.experts.{base_layer}w13_weight", ".moe.up_proj.weight", "w3"),
+            (f".moe.experts.{base_layer}w2_weight", ".moe.down_proj.weight", "w2"),
         ]
 
         # New per-expert format: .moe.experts.E.gate_proj.weight_packed [out, in]
diff --git a/vllm/model_executor/models/step3p5_mtp.py b/vllm/model_executor/models/step3p5_mtp.py
index 83e43dce511..092f7a31aa7 100644
--- a/vllm/model_executor/models/step3p5_mtp.py
+++ b/vllm/model_executor/models/step3p5_mtp.py
@@ -181,14 +181,17 @@ class Step3p5MTP(nn.Module):
             ("gate_up_proj", "gate_proj", 0),
             ("gate_up_proj", "up_proj", 1),
         ]
+        params_dict = dict(self.named_parameters())
+        base_layer = (
+            "base_layer." if any(".base_layer." in name for name in params_dict) else ""
+        )
 
         expert_params_mapping = [
-            (".moe.experts.w13_weight", ".moe.gate_proj.weight", "w1"),
-            (".moe.experts.w13_weight", ".moe.up_proj.weight", "w3"),
-            (".moe.experts.w2_weight", ".moe.down_proj.weight", "w2"),
+            (f".moe.experts.{base_layer}w13_weight", ".moe.gate_proj.weight", "w1"),
+            (f".moe.experts.{base_layer}w13_weight", ".moe.up_proj.weight", "w3"),
+            (f".moe.experts.{base_layer}w2_weight", ".moe.down_proj.weight", "w2"),
         ]
 
-        params_dict = dict(self.named_parameters())
         loaded_params: set[str] = set()
         for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name: