diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index e54fa057..86e78fbd 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -187,6 +187,7 @@ target_link_libraries(flm PUBLIC
     phi4_npu
     nanbeige_npu
     dequant
+    dequant_new
     gemm
     lm_head
     mha
diff --git a/src/include/modules/dequant.hpp b/src/include/modules/dequant.hpp
index 43596d08..f1c236ac 100644
--- a/src/include/modules/dequant.hpp
+++ b/src/include/modules/dequant.hpp
@@ -21,12 +21,17 @@ class Dequant{
     Dequant(LM_Config& config);
     ~Dequant();
 
-    /// @brief generate the dequant sequence
-    /// @param seq: the sequence
-    /// @param D_in: input dimension of the projection weight
-    /// @param D_out: output dimension of the projection weight
-    /// @param weight_offset: the weight offset in byte
-    /// @param mode: dequant output mode
+    typedef enum: int {
+        Q4_1 = 0,
+        Q8_0 = 1
+    } quant_block_t;
+
+    void reorder_cpy(u8 *dst, buffer<u8> &src,
+        quant_block_t quant_block_type,
+        const int quant_matrix_row,
+        const int quant_matrix_col,
+        const int vertical_blocks=2,
+        const int vetrical_block_interleave_byte_size=-1);
     void generate_dequant_q4_1_seq(npu_sequence* seq, const uint32_t D_in, const uint32_t D_out, const uint32_t weight_offset, int mode);
     void generate_dequant_q80_packed_in_q4nx_seq(npu_sequence* seq, const uint32_t D_in, const uint32_t D_out, const uint32_t weight_offset, int mode);
 private:
diff --git a/src/lib/libdequant_new.so b/src/lib/libdequant_new.so
new file mode 100755
index 00000000..fc02a73a
Binary files /dev/null and b/src/lib/libdequant_new.so differ
diff --git a/src/lib/libllama_npu.so b/src/lib/libllama_npu.so
index ef423e72..b5c3ce29 100755
Binary files a/src/lib/libllama_npu.so and b/src/lib/libllama_npu.so differ
diff --git a/src/lib/libphi4_npu.so b/src/lib/libphi4_npu.so
index 13cd76c2..40065d39 100755
Binary files a/src/lib/libphi4_npu.so and b/src/lib/libphi4_npu.so differ
diff --git a/src/lib/libq4_npu_eXpress.so b/src/lib/libq4_npu_eXpress.so
index f0d0bd6b..5dd4e8f5 100755
Binary files a/src/lib/libq4_npu_eXpress.so and b/src/lib/libq4_npu_eXpress.so differ
diff --git a/src/lib/libqwen3_5vl_npu.so b/src/lib/libqwen3_5vl_npu.so
index f1aa361e..837d87ee 100755
Binary files a/src/lib/libqwen3_5vl_npu.so and b/src/lib/libqwen3_5vl_npu.so differ
diff --git a/src/lib/libqwen3_npu.so b/src/lib/libqwen3_npu.so
index f023f1f0..14db01d4 100755
Binary files a/src/lib/libqwen3_npu.so and b/src/lib/libqwen3_npu.so differ
diff --git a/src/lib/libqwen3vl_npu.so b/src/lib/libqwen3vl_npu.so
index 19e9dc3d..ebd7f1c4 100755
Binary files a/src/lib/libqwen3vl_npu.so and b/src/lib/libqwen3vl_npu.so differ
diff --git a/src/test/CMakeLists.txt b/src/test/CMakeLists.txt
index 8d925855..5873e8f9 100644
--- a/src/test/CMakeLists.txt
+++ b/src/test/CMakeLists.txt
@@ -110,6 +110,7 @@ function(add_npu_test TARGET_NAME OUTPUT_SUBDIR)
         q4_npu_eXpress
         lm_head
         dequant
+        dequant_new
         gemm
         mha
         tokenizers_cpp
diff --git a/src/xclbins/Qwen3.5-0.8B-NPU2/dequant_q41.xclbin b/src/xclbins/Qwen3.5-0.8B-NPU2/dequant_q41.xclbin
new file mode 100644
index 00000000..18e227f5
Binary files /dev/null and b/src/xclbins/Qwen3.5-0.8B-NPU2/dequant_q41.xclbin differ
diff --git a/src/xclbins/Qwen3.5-0.8B-NPU2/dequant_q80.xclbin b/src/xclbins/Qwen3.5-0.8B-NPU2/dequant_q80.xclbin
new file mode 100644
index 00000000..e3143417
Binary files /dev/null and b/src/xclbins/Qwen3.5-0.8B-NPU2/dequant_q80.xclbin differ
diff --git a/src/xclbins/Qwen3.5-0.8B-NPU2/layer.xclbin b/src/xclbins/Qwen3.5-0.8B-NPU2/layer.xclbin
index 909d5432..9ad7a95a 100644
Binary files a/src/xclbins/Qwen3.5-0.8B-NPU2/layer.xclbin and b/src/xclbins/Qwen3.5-0.8B-NPU2/layer.xclbin differ
diff --git a/src/xclbins/Qwen3.5-0.8B-NPU2/lm_head.xclbin b/src/xclbins/Qwen3.5-0.8B-NPU2/lm_head.xclbin
index dc119b32..e645706e 100644
Binary files a/src/xclbins/Qwen3.5-0.8B-NPU2/lm_head.xclbin and b/src/xclbins/Qwen3.5-0.8B-NPU2/lm_head.xclbin differ