diff --git a/src/plugins/intel_cpu/src/transformations/cpu_opset/convert_to_cpu_specific_opset.hpp b/src/plugins/intel_cpu/src/transformations/cpu_opset/convert_to_cpu_specific_opset.hpp
index e334d11babe0ee..4139d5c6cf8042 100644
--- a/src/plugins/intel_cpu/src/transformations/cpu_opset/convert_to_cpu_specific_opset.hpp
+++ b/src/plugins/intel_cpu/src/transformations/cpu_opset/convert_to_cpu_specific_opset.hpp
@@ -15,11 +15,9 @@
 #include "common/pass/convert_to_swish_cpu.hpp"
 #include "common/pass/move_fc_reshape_to_weights.hpp"
 #include "transformations/convert_precision.hpp"
-#include "transformations/symbolic_transformations/symbolic_optimizations.hpp"
 #include "transformations/utils/utils.hpp"
 #include "common/pass/rnn_sequences_optimization.hpp"
 #include "transformations/common_optimizations/reshape_sequence_fusion.hpp"
-#include "common/pass/ngram_fusion.hpp"
 #include "transformations/defs.hpp"
 
 #include "itt.hpp"
@@ -50,8 +48,6 @@ inline void ConvertToCPUSpecificOpset(std::shared_ptr<ov::Model> &nGraphFunc) {
                              type_to_fuse_map{{}},
                              false,
                              false);
-    auto symbolic_pipeline = CPU_REGISTER_PASS_COMMON(manager, ov::pass::SymbolicOptimizations, false);
-    symbolic_pipeline->get_manager()->register_pass<NgramFusion>();
     CPU_REGISTER_PASS_COMMON(manager, ov::pass::Validate);
 
     manager.run_passes(nGraphFunc);
diff --git a/src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp b/src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp
index a4522436fbbaaf..817f346d670c5f 100644
--- a/src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp
+++ b/src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp
@@ -83,6 +83,7 @@
 #include "transformations/opset_conversions/convert_opset2_to_opset1.hpp"
 #include "transformations/opset_conversions/convert_opset3_to_opset2.hpp"
 #include "transformations/smart_reshape/matmul_sr.hpp"
+#include "transformations/symbolic_transformations/symbolic_optimizations.hpp"
 #include "transformations/init_node_info.hpp"
 #include "utils/ngraph_transformation.hpp"
 #include "utils/print_model.hpp"
@@ -113,6 +114,7 @@
 #include "transformations/cpu_opset/common/pass/convert_fq_rnn_to_quantized_rnn.hpp"
 #include "transformations/cpu_opset/common/pass/insert_convert_after_extension.hpp"
 #include "transformations/cpu_opset/common/pass/move_eltwise_up_data_movement.hpp"
+#include "transformations/cpu_opset/common/pass/ngram_fusion.hpp"
 #include "transformations/cpu_opset/common/pass/permute_slice_n_interpolation.hpp"
 #include "transformations/cpu_opset/common/pass/swap_convert_transpose.hpp"
 #include "transformations/cpu_opset/common/pass/rope_fusion.hpp"
@@ -708,6 +710,11 @@ void Transformations::PostLpt() {
     CPU_REGISTER_PASS_X64(postLPTPassManager, RoPEFusion);
 
     CPU_REGISTER_PASS_X64(postLPTPassManager, StatefulSDPAFusion);
+
+    // Should be before Snippets pipeline because Ngram pattern contains eltwise nodes that can be tokenized by Snippets.
+    auto symbolic_pipeline = CPU_REGISTER_PASS_COMMON(postLPTPassManager, ov::pass::SymbolicOptimizations, false);
+    symbolic_pipeline->get_manager()->register_pass<NgramFusion>();
+
     postLPTPassManager.run_passes(model);
 }