diff --git a/ggml/src/ggml-openvino/ggml-decoder.h b/ggml/src/ggml-openvino/ggml-decoder.h
index 8e39a26c8b79..91b14211c270 100644
--- a/ggml/src/ggml-openvino/ggml-decoder.h
+++ b/ggml/src/ggml-openvino/ggml-decoder.h
@@ -23,7 +23,7 @@ struct ModelParams {
     int n_heads_kv = -1;
     int head_size = -1;
     int state_size = -1;  // for SSM molels, eg qwen35
-    int32_t rope_params[15];
+    int32_t rope_params[15] = {0};
     bool mixed_rope_params = false;
     std::vector<int> swa_layers;
 
diff --git a/ggml/src/ggml-openvino/openvino/pass/fuse_to_conv.cpp b/ggml/src/ggml-openvino/openvino/pass/fuse_to_conv.cpp
new file mode 100644
index 000000000000..21801c0f3992
--- /dev/null
+++ b/ggml/src/ggml-openvino/openvino/pass/fuse_to_conv.cpp
@@ -0,0 +1,212 @@
+#include "fuse_to_conv.h"
+
+#include <openvino/core/graph_util.hpp>
+#include <openvino/core/rt_info.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/convolution.hpp>
+#include <openvino/op/extractimagepatches.hpp>
+#include <openvino/op/matmul.hpp>
+#include <openvino/op/pad.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/transpose.hpp>
+#include <openvino/pass/pattern/op/label.hpp>
+#include <openvino/pass/pattern/op/pattern.hpp>
+#include <openvino/pass/pattern/op/wrap_type.hpp>
+
+namespace opp = ov::pass::pattern;
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+// This pass fuses an IM2COL + MatMul convolution into OpenVINO's Convolution op for performance gains.
+// Reference the im2col.cpp translator for reference on the pattern being matched.
+
+FuseToConv::FuseToConv() {
+    const auto m_wei = opp::any_input();
+    const auto m_act = opp::any_input();
+    const auto m_matmul = opp::wrap_type<ov::op::v0::MatMul>({m_wei, m_act});
+
+    const auto callback = [=](ov::pass::pattern::Matcher & m) {
+        const auto & pm = m.get_pattern_value_map();
+
+        auto matmul_node = ov::as_type_ptr<ov::op::v0::MatMul>(pm.at(m_matmul).get_node_shared_ptr());
+        if (!matmul_node || matmul_node->get_transpose_a() || !matmul_node->get_transpose_b()) {
+            return false;
+        }
+
+        auto trace = matmul_node->input_value(1);
+
+        // Optional Convert
+        if (auto n = ov::as_type_ptr<ov::op::v0::Convert>(trace.get_node_shared_ptr())) {
+            trace = n->input_value(0);
+        }
+
+        for (int i = 0; i < 2; ++i) {
+            auto n = ov::as_type_ptr<ov::op::v1::Reshape>(trace.get_node_shared_ptr());
+            if (!n) {
+                return false;
+            }
+            trace = n->input_value(0);
+        }
+
+        if (auto n = ov::as_type_ptr<ov::op::v1::Transpose>(trace.get_node_shared_ptr())) {
+            trace = n->input_value(0);
+        } else {
+            return false;
+        }
+
+        if (auto n = ov::as_type_ptr<ov::op::v1::Reshape>(trace.get_node_shared_ptr())) {
+            trace = n->input_value(0);
+        } else {
+            return false;
+        }
+
+        if (auto n = ov::as_type_ptr<ov::op::v1::Transpose>(trace.get_node_shared_ptr())) {
+            trace = n->input_value(0);
+        } else {
+            return false;
+        }
+
+        auto eip = ov::as_type_ptr<ov::op::v3::ExtractImagePatches>(trace.get_node_shared_ptr());
+        if (!eip) {
+            return false;
+        }
+        const auto eip_strides = eip->get_strides();  // {stride_h, stride_w}
+        const auto eip_rates = eip->get_rates();      // {dil_h, dil_w}
+
+        auto pad = ov::as_type_ptr<ov::op::v1::Pad>(eip->input_value(0).get_node_shared_ptr());
+        if (!pad) {
+            return false;
+        }
+        auto pads_begin_const =
+            ov::as_type_ptr<ov::op::v0::Constant>(pad->input_value(1).get_node_shared_ptr());
+
+        const auto pads_begin_vals = pads_begin_const->cast_vector<int64_t>();  // {0, 0, pad_h, pad_w}
+        const std::ptrdiff_t pad_h = static_cast<std::ptrdiff_t>(pads_begin_vals[2]);
+        const std::ptrdiff_t pad_w = static_cast<std::ptrdiff_t>(pads_begin_vals[3]);
+
+        auto image_input = pad->input_value(0);  // [N, IC, 1, IW] NCHW
+
+        auto w_trace = matmul_node->input_value(0);
+        if (auto n = ov::as_type_ptr<ov::op::v0::Convert>(w_trace.get_node_shared_ptr())) {
+            w_trace = n->input_value(0);
+        }
+        for (int i = 0; i < 2; ++i) {
+            auto n = ov::as_type_ptr<ov::op::v1::Reshape>(w_trace.get_node_shared_ptr());
+            if (!n) {
+                break;
+            }
+            w_trace = n->input_value(0);
+        }
+
+        auto weight_const = ov::as_type_ptr<ov::op::v0::Constant>(w_trace.get_node_shared_ptr());
+        if (!weight_const) {
+            return false;
+        }
+
+        // Reshape weight to [OC, IC, 1, KW] (OIHW).
+        const auto w_shape = weight_const->get_shape();
+        ov::Shape conv_w_shape;
+        if (w_shape.size() == 3) {
+            conv_w_shape = {w_shape[0], w_shape[1], 1, w_shape[2]};
+        } else if (w_shape.size() == 4) {
+            conv_w_shape = {w_shape[1], w_shape[2], 1, w_shape[3]};
+        } else {
+            return false;
+        }
+
+        auto weight_reshaped = register_new_node<ov::op::v0::Constant>(weight_const->get_element_type(), conv_w_shape,
+                                                                       weight_const->get_data_ptr());
+
+        ov::Output<Node> weight_input = weight_reshaped;
+        if (weight_reshaped->get_element_type() != image_input.get_element_type()) {
+            weight_input = register_new_node<ov::op::v0::Convert>(weight_reshaped, image_input.get_element_type());
+        }
+
+        auto conv = register_new_node<ov::op::v1::Convolution>(
+            image_input, weight_input,
+            ov::Strides{static_cast<size_t>(eip_strides[0]), static_cast<size_t>(eip_strides[1])},
+            ov::CoordinateDiff{pad_h, pad_w}, ov::CoordinateDiff{pad_h, pad_w},
+            ov::Strides{static_cast<size_t>(eip_rates[0]), static_cast<size_t>(eip_rates[1])},
+            ov::op::PadType::EXPLICIT);
+
+        constexpr auto target_type = ov::element::f32;
+        ov::Output<Node> conv_out = conv;
+        if (conv_out.get_element_type() != target_type) {
+            conv_out = register_new_node<ov::op::v0::Convert>(conv_out, target_type);
+        }
+
+        std::shared_ptr<ov::op::v1::Add> add_node;
+        ov::Output<Node> bias_input;
+        for (const auto & consumer_in : matmul_node->output(0).get_target_inputs()) {
+            auto cast = ov::as_type_ptr<ov::op::v0::Convert>(consumer_in.get_node()->shared_from_this());
+            if (!cast) {
+                continue;
+            }
+            for (const auto & add_in : cast->output(0).get_target_inputs()) {
+                auto add = ov::as_type_ptr<ov::op::v1::Add>(add_in.get_node()->shared_from_this());
+                if (!add) {
+                    continue;
+                }
+                for (size_t i = 0; i < 2; ++i) {
+                    if (ov::as_type_ptr<ov::op::v0::Constant>(add->input_value(i).get_node_shared_ptr())) {
+                        bias_input = add->input_value(i);
+                        add_node = add;
+                        break;
+                    }
+                }
+                if (add_node) {
+                    break;
+                }
+            }
+            if (add_node) {
+                break;
+            }
+        }
+
+        ov::Output<Node> final_out;
+        std::shared_ptr<Node> target_node;
+
+        if (add_node) {
+            // Reshape bias [OC, 1] → [1, OC, 1, 1] for NCHW broadcasting.
+            ov::Output<Node> bias = bias_input;
+            if (bias.get_element_type() != target_type) {
+                bias = register_new_node<ov::op::v0::Convert>(bias, target_type);
+            }
+            const auto oc = static_cast<int64_t>(conv_w_shape[0]);
+            auto bias_shape = register_new_node<ov::op::v0::Constant>(ov::element::i64, ov::Shape{4},
+                                                                      std::vector<int64_t>{1, oc, 1, 1});
+            bias = register_new_node<ov::op::v1::Reshape>(bias, bias_shape, false);
+            final_out = register_new_node<ov::op::v1::Add>(conv_out, bias);
+            target_node = add_node;
+        } else {
+            final_out = conv_out;
+            target_node = matmul_node;
+        }
+
+        // Reshape final output back to the target node's original shape if needed.
+        auto orig_shape = target_node->get_output_partial_shape(0);
+        if (orig_shape.is_static() && final_out.get_partial_shape() != orig_shape) {
+            auto shape_const = register_new_node<ov::op::v0::Constant>(ov::element::i64, ov::Shape{orig_shape.size()},
+                                                                       orig_shape.to_shape());
+            final_out = register_new_node<ov::op::v1::Reshape>(final_out, shape_const, false);
+        }
+
+        final_out.get_node_shared_ptr()->set_friendly_name(target_node->get_friendly_name());
+        ov::copy_runtime_info(m.get_matched_nodes(), final_out.get_node_shared_ptr());
+        ov::replace_node(target_node, final_out.get_node_shared_ptr());
+
+        return true;
+    };
+
+    register_matcher(std::make_shared<opp::Matcher>(m_matmul, "ov::frontend::ggml::pass::FuseToConv"), callback);
+}
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov
diff --git a/ggml/src/ggml-openvino/openvino/pass/fuse_to_conv.h b/ggml/src/ggml-openvino/openvino/pass/fuse_to_conv.h
new file mode 100644
index 000000000000..feac14b13ff2
--- /dev/null
+++ b/ggml/src/ggml-openvino/openvino/pass/fuse_to_conv.h
@@ -0,0 +1,17 @@
+#include "openvino/pass/matcher_pass.hpp"
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+class FuseToConv : public ov::pass::MatcherPass {
+public:
+    OPENVINO_MATCHER_PASS_RTTI("ov::frontend::ggml::pass::FuseToConv")
+    FuseToConv();
+};
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov
diff --git a/ggml/src/ggml-openvino/openvino/translate_session.cpp b/ggml/src/ggml-openvino/openvino/translate_session.cpp
index 35598aba6be8..da8b543f5d4a 100644
--- a/ggml/src/ggml-openvino/openvino/translate_session.cpp
+++ b/ggml/src/ggml-openvino/openvino/translate_session.cpp
@@ -5,6 +5,7 @@
 #include "ggml-openvino/openvino/node_context.h"
 #include "ggml-openvino/openvino/utils.h"
 #include "input_model.h"
+#include "pass/fuse_to_conv.h"
 #include "pass/mark_decompression_convert_constant_folding.h"
 #include "pass/mark_dequantization_subgraph.h"
 #include "pass/squeeze_matmul.h"
@@ -395,6 +396,7 @@ std::shared_ptr<Model> TranslateSession::apply_transformations(std::shared_ptr<M
         // is_decompression_multiply() recognizes GatherMatmul as a valid consumer.
         manager.register_pass<ov::pass::MarkDequantization>(
             std::vector<ov::element::Type>{ov::element::u8, ov::element::i8, ov::element::u4, ov::element::i4});
+        manager.register_pass<pass::FuseToConv>();
 
         if (ggml_model_decoder->is_stateful()) {
             const auto kv_param_res_names = ggml_model_decoder->get_kv_param_res_names();
diff --git a/ggml/src/ggml-openvino/utils.h b/ggml/src/ggml-openvino/utils.h
index 513fa83c9d6e..3a7e4fa8c7e2 100644
--- a/ggml/src/ggml-openvino/utils.h
+++ b/ggml/src/ggml-openvino/utils.h
@@ -1,9 +1,9 @@
 #include "ggml-decoder.h"
 #include "ggml-impl.h"
 
-#include <algorithm>
 #include <atomic>
 #include <cstddef>
+#include <cstdint>
 #include <functional>
 #include <memory>
 #include <mutex>
@@ -11,6 +11,7 @@
 #include <openvino/runtime/infer_request.hpp>
 #include <string>
 #include <unordered_map>
+#include <unordered_set>
 #include <utility>
 #include <vector>
 
@@ -18,7 +19,7 @@ struct graph_key {
     int n_nodes;
     std::string first_node_name;
     std::string last_node_name;
-    std::vector<std::string> input_src_names;
+    std::vector<std::string> input_src_buffers;
 
     graph_key(const ggml_cgraph * cgraph) : n_nodes(cgraph->n_nodes) {
         if (n_nodes > 0) {
@@ -26,46 +27,33 @@ struct graph_key {
             last_node_name = cgraph->nodes[n_nodes - 1]->name;
         }
 
-        auto get_input_key_name = [](const ggml_cgraph * graph, const ggml_tensor * tensor) {
-            std::string name = tensor->name;
-            const size_t hash_pos = ggml_hash_find(&graph->visited_hash_set, tensor);
-            if (((tensor->flags & GGML_TENSOR_FLAG_COMPUTE) || GgmlOvDecoder::is_kvcache(tensor, nullptr)) &&
-                hash_pos != GGML_HASHSET_FULL && ggml_bitset_get(graph->visited_hash_set.used, hash_pos)) {
-                name += "#" + std::to_string(hash_pos);
-            }
-            return name;
-        };
-
-        std::vector<std::string> node_names;
-        node_names.reserve(cgraph->n_nodes);
-        for (int node_idx = 0; node_idx < cgraph->n_nodes; node_idx++) {
-            node_names.emplace_back(cgraph->nodes[node_idx]->name);
+        std::unordered_set<const ggml_tensor *> node_set;
+        node_set.reserve(cgraph->n_nodes);
+        for (int i = 0; i < cgraph->n_nodes; i++) {
+            node_set.insert(cgraph->nodes[i]);
         }
 
         for (int node_idx = 0; node_idx < cgraph->n_nodes; node_idx++) {
             const ggml_tensor * node = cgraph->nodes[node_idx];
             for (int src_idx = 0; src_idx < GGML_MAX_SRC; src_idx++) {
                 const ggml_tensor * src = node->src[src_idx];
-                if (src == nullptr || src->name[0] == '\0') {
+                if (src == nullptr || src->buffer == nullptr || node_set.count(src) || src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
                     continue;
                 }
 
-                const std::string src_name = get_input_key_name(cgraph, src);
-                if (std::find(node_names.begin(), node_names.end(), src_name) != node_names.end()) {
-                    continue;
-                }
-                if (src_name.find("weight") != std::string::npos) {
-                    continue;
-                }
+                const uintptr_t buf_ptr = reinterpret_cast<uintptr_t>(src->buffer);
+                const uintptr_t base = reinterpret_cast<uintptr_t>(ggml_backend_buffer_get_base(src->buffer));
+                const uintptr_t offset = reinterpret_cast<uintptr_t>(src->data) - base;
 
-                input_src_names.push_back(std::to_string(node_idx) + ":" + std::to_string(src_idx) + ":" + src_name);
+                input_src_buffers.push_back(std::to_string(node_idx) + ":" + std::to_string(src_idx) + ":" +
+                    std::to_string(buf_ptr) + "+" + std::to_string(offset));
             }
         }
     }
 
     bool operator==(const graph_key & other) const {
         return n_nodes == other.n_nodes && first_node_name == other.first_node_name &&
-               last_node_name == other.last_node_name && input_src_names == other.input_src_names;
+               last_node_name == other.last_node_name && input_src_buffers == other.input_src_buffers;
     }
 };
 
@@ -76,8 +64,8 @@ struct graph_key_hash {
             hash ^= std::hash<std::string>{}(key.first_node_name) + 0x9e3779b9 + (hash << 6) + (hash >> 2);
             hash ^= std::hash<std::string>{}(key.last_node_name) + 0x9e3779b9 + (hash << 6) + (hash >> 2);
         }
-        for (const auto & input_src_name : key.input_src_names) {
-            hash ^= std::hash<std::string>{}(input_src_name) + 0x9e3779b9 + (hash << 6) + (hash >> 2);
+        for (const auto & input_src_buffer : key.input_src_buffers) {
+            hash ^= std::hash<std::string>{}(input_src_buffer) + 0x9e3779b9 + (hash << 6) + (hash >> 2);
         }
         return hash;
     }