diff --git a/vllm/model_executor/layers/quantization/moe_wna16.py b/vllm/model_executor/layers/quantization/moe_wna16.py
index e5ef3f4c3..e27f28c06 100644
--- a/vllm/model_executor/layers/quantization/moe_wna16.py
+++ b/vllm/model_executor/layers/quantization/moe_wna16.py
@@ -1,11 +1,13 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+import os
 from typing import Any
 
 import torch
 
 from vllm.distributed import get_tensor_model_parallel_rank, get_tp_group
+from vllm.logger import init_logger
 from vllm.model_executor.layers.fused_moe.activation import MoEActivation
 from vllm.model_executor.layers.fused_moe.config import (
     FusedMoEQuantConfig,
@@ -33,6 +35,8 @@ from vllm.model_executor.layers.quantization.utils.marlin_utils import (
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.platforms import current_platform
 
+logger = init_logger(__name__)
+
 
 class MoeWNA16Config(QuantizationConfig):
     """Config class for MOE WNA16 (W8A16/W4A16) quantization."""
@@ -362,6 +366,39 @@ class MoeWNA16Method(FusedMoEMethodBase):
             block_shape=[0, layer.group_size],
         )
 
+    def _llm_scaler_moe_requested(self) -> bool:
+        return os.environ.get("VLLM_XPU_USE_LLM_SCALER_MOE", "0") == "1"
+
+    def _can_use_llm_scaler_moe(self, layer: torch.nn.Module) -> bool:
+        return (
+            self._llm_scaler_moe_requested()
+            and current_platform.is_xpu()
+            and self.quant_config.weight_bits == 4
+            and not self.quant_config.has_zp
+            and getattr(layer, "expert_map", None) is None
+            and getattr(layer, "global_num_experts", None)
+            == getattr(layer, "local_num_experts", None)
+            and not getattr(layer, "apply_router_weight_on_input", False)
+            and getattr(layer, "group_size", None) == 128
+            and getattr(layer, "w13_qweight", None) is not None
+            and getattr(layer, "w2_qweight", None) is not None
+            and layer.w13_qweight.dtype == torch.uint8
+            and layer.w2_qweight.dtype == torch.uint8
+            and layer.w13_scales.dtype == torch.float16
+            and layer.w2_scales.dtype == torch.float16
+        )
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        if not self._can_use_llm_scaler_moe(layer):
+            return
+        if getattr(layer, "_llm_scaler_moe_u4_decode", False):
+            return
+
+        layer._llm_scaler_moe_u4_decode = True
+        logger.info_once(
+            "Enabled llm-scaler XPU INT4 MoE decode path for %s", layer.layer_name
+        )
+
     def apply(
         self,
         layer: FusedMoE,
@@ -376,6 +413,30 @@ class MoeWNA16Method(FusedMoEMethodBase):
             f"Only SiLU activation is supported, not {layer.activation}."
         )
 
+        if (
+            getattr(layer, "_llm_scaler_moe_u4_decode", False)
+            and shared_experts_input is None
+            and x.dtype == torch.float16
+            and x.shape[0] <= 4
+        ):
+            try:
+                from custom_esimd_kernels_vllm import (
+                    moe_forward_tiny_cutlass_nmajor_int4_u4,
+                )
+
+                return moe_forward_tiny_cutlass_nmajor_int4_u4(
+                    x,
+                    layer.w13_qweight,
+                    layer.w13_scales,
+                    layer.w2_qweight,
+                    layer.w2_scales,
+                    topk_weights.contiguous(),
+                    topk_ids.contiguous(),
+                )
+            except Exception:
+                logger.exception("llm-scaler XPU INT4 MoE path failed")
+                raise
+
         return fused_experts(
             x,
             layer.w13_qweight,