diff --git a/mediapipe/BUILD b/mediapipe/BUILD
index 0810f23582c..51c91987c58 100644
--- a/mediapipe/BUILD
+++ b/mediapipe/BUILD
@@ -215,8 +215,6 @@ filegroup(
   srcs = [
     "//mediapipe/calculators/geti/serialization:calculators",
     "//mediapipe/calculators/geti/inference:geti_calculator_base",
-    "//mediapipe/calculators/geti/inference:grpc_predict_v2_cc_proto",
-    "//mediapipe/calculators/geti/inference:grpc_predict_v2_proto",
     "//mediapipe/calculators/geti/inference:inference_calculators",
     "//mediapipe/calculators/geti/inference:inference_utils",
     "//mediapipe/calculators/geti/inference:openvino_inference_adapter_calculator_cc_proto",
diff --git a/mediapipe/calculators/geti/inference/BUILD b/mediapipe/calculators/geti/inference/BUILD
index c7bada606fe..1bf2c10cd44 100644
--- a/mediapipe/calculators/geti/inference/BUILD
+++ b/mediapipe/calculators/geti/inference/BUILD
@@ -1,18 +1,7 @@
 load("//mediapipe/framework/port:build_config.bzl", "mediapipe_cc_proto_library", "mediapipe_proto_library")
-load("@rules_cc//cc:defs.bzl", "cc_proto_library")
 
 package(default_visibility = ["//visibility:public"])
 
-proto_library(
-    name = "grpc_predict_v2_proto",
-    srcs = ["grpc_predict_v2.proto"],
-)
-
-cc_proto_library(
-    name = "grpc_predict_v2_cc_proto",
-    deps = [":grpc_predict_v2_proto"],
-)
-
 proto_library(
     name = "openvino_inference_adapter_calculator_proto",
     srcs = ["openvino_inference_adapter_calculator.proto"],
@@ -63,7 +52,7 @@ cc_library(
     ],
     deps = [
         ":openvino_inference_adapter_calculator_cc_proto",
-        ":grpc_predict_v2_cc_proto",
+        "@ovms//src/kfserving_api:kfserving_api_cpp",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/formats:image_frame_opencv",
         "//mediapipe/framework/port:opencv_core",
diff --git a/mediapipe/calculators/geti/inference/kserve.h b/mediapipe/calculators/geti/inference/kserve.h
index f5db488d68c..811912ce69c 100644
--- a/mediapipe/calculators/geti/inference/kserve.h
+++ b/mediapipe/calculators/geti/inference/kserve.h
@@ -16,7 +16,7 @@
 #ifndef KSERVE_H
 #define KSERVE_H
 
-#include "mediapipe/calculators/geti/inference/grpc_predict_v2.pb.h"
+#include "src/kfserving_api/grpc_predict_v2.pb.h"
 
 using KFSRequest = inference::ModelInferRequest;
 using KFSResponse = inference::ModelInferResponse;
diff --git a/mediapipe/calculators/ovms/BUILD b/mediapipe/calculators/ovms/BUILD
index b32f05e5e1e..5efa8e8a6fa 100644
--- a/mediapipe/calculators/ovms/BUILD
+++ b/mediapipe/calculators/ovms/BUILD
@@ -38,6 +38,26 @@ cc_library(
     alwayslink = 1,
 )
 
+cc_library(
+    name = "ovms_calculator_runtime",
+    srcs = [
+    ],
+    deps = [
+        "//mediapipe/calculators/openvino:openvino_tensors_to_classification_calculator_cc_proto",
+        "//mediapipe/calculators/openvino:openvino_tensors_to_detections_calculator_cc_proto",
+        "//mediapipe/calculators/openvino:openvino_converter_calculator_cc_proto",
+        "//mediapipe/calculators/openvino:openvino_converter_calculator",
+        "//mediapipe/calculators/openvino:openvino_tensors_to_classification_calculator",
+        "//mediapipe/calculators/openvino:openvino_tensors_to_detections_calculator",
+        ":modelapiovmsadapter",
+        ":openvinoinferencecalculator_no_tflite",
+        ":openvinomodelserversessioncalculator",
+    ],
+    copts = ["-Iexternal/ovms/src","-Isrc"],
+    linkopts = ["-Lmediapipe/"],
+    alwayslink = 1,
+)
+
 cc_library(
     name = "modelapiovmsadapter",
     srcs = [
@@ -125,6 +145,28 @@ cc_library(
     alwayslink = 1,
 )
 
+cc_library(
+    name = "openvinoinferencecalculator_no_tflite",
+    srcs = [
+        "openvinoinferencecalculator.cc",
+    ],
+    hdrs = [
+        "openvinoinferencecalculator.h",
+    ],
+    deps = [
+        "openvinoinferencecalculatoroptions",
+        ":modelapiovmsadapter",
+        ":openvinoinferencecalculator_cc_proto",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:tensor",
+        "//third_party:openvino",
+        "@ovms//src:ovms_header",
+    ],
+    copts = ["-Iexternal/ovms/src","-Isrc"],
+    local_defines = ["OVMS_RUNTIME_DISABLE_TFLITE_TENSORS=1"],
+    alwayslink = 1,
+)
+
 cc_library(
     name = "openvinomodelserversessioncalculator",
     srcs = [
diff --git a/mediapipe/calculators/ovms/modelapiovmsadapter.cc b/mediapipe/calculators/ovms/modelapiovmsadapter.cc
index 8065ace395d..87680ab7033 100644
--- a/mediapipe/calculators/ovms/modelapiovmsadapter.cc
+++ b/mediapipe/calculators/ovms/modelapiovmsadapter.cc
@@ -36,7 +36,6 @@
 // for the one inside OVMS repo it makes sense to reuse code from ovms lib
 namespace mediapipe::ovms {
     
-using std::endl;
 #pragma GCC diagnostic push
 #pragma GCC diagnostic ignored "-Wunused-function"
 using InferenceOutput = std::map<std::string, ov::Tensor>;
diff --git a/mediapipe/calculators/ovms/modelapiovmsadapter.hpp b/mediapipe/calculators/ovms/modelapiovmsadapter.hpp
index 7a315bd9505..9d78c7c8967 100644
--- a/mediapipe/calculators/ovms/modelapiovmsadapter.hpp
+++ b/mediapipe/calculators/ovms/modelapiovmsadapter.hpp
@@ -17,8 +17,12 @@
 #include <iostream>
 #include <map>
 #include <memory>
+#include <mutex>
+#include <cstdio>
 #include <sstream>
+#include <stdexcept>
 #include <string>
+#include <thread>
 #include <unordered_map>
 #include <utility>
 #include <vector>
@@ -28,6 +32,10 @@
 
 #include "ovms.h"  // NOLINT
 
+#ifdef _WIN32
+#include <Windows.h>
+#endif
+
 // here we need to decide if we have several calculators (1 for OVMS repository, 1-N inside mediapipe)
 // for the one inside OVMS repo it makes sense to reuse code from ovms lib
 
@@ -54,6 +62,45 @@ class OVMSInferenceAdapter : public ::InferenceAdapter {
     std::unordered_map<std::string, ov::element::Type_t> outputDatatypes;
     ov::AnyMap modelConfig;
 
+    // Creates (once per process) and returns a shared OVMS server handle for adapters
+    // that are constructed without an explicit server pointer.
+    // This path is used by calculators running in-process when they rely on the default
+    // OVMS singleton instead of receiving a handle from runtime/shared-library plumbing.
+    // Keeping one shared handle avoids repeated OVMS_ServerNew calls and keeps all such
+    // adapters bound to the same server instance.
+    static OVMS_Server* getSharedServerHandle() {
+        static std::once_flag once;
+        static OVMS_Server* sharedServer{nullptr};
+        std::fprintf(stderr, "OVMSAdapter shared handle call_once entry thread=%zu\n", std::hash<std::thread::id>{}(std::this_thread::get_id()));
+        std::call_once(once, []() {
+            std::fprintf(stderr, "OVMSAdapter shared handle init start\n");
+            auto* serverNewFn = &OVMS_ServerNew;
+            std::fprintf(stderr, "OVMSAdapter shared handle OVMS_ServerNew fn=%p\n", reinterpret_cast<void*>(serverNewFn));
+#ifdef _WIN32
+            HMODULE mod = nullptr;
+            if (GetModuleHandleExA(GET_MODULE_HANDLE_EX_FLAG_FROM_ADDRESS | GET_MODULE_HANDLE_EX_FLAG_UNCHANGED_REFCOUNT,
+                    reinterpret_cast<LPCSTR>(serverNewFn), &mod) != 0) {
+                char modulePath[MAX_PATH] = {0};
+                DWORD pathLen = GetModuleFileNameA(mod, modulePath, MAX_PATH);
+                if (pathLen > 0) {
+                    std::fprintf(stderr, "OVMSAdapter shared handle OVMS_ServerNew module=%s\n", modulePath);
+                }
+            }
+#endif
+            OVMS_Status* status = serverNewFn(&sharedServer);
+            std::fprintf(stderr, "OVMSAdapter shared handle init OVMS_ServerNew returned status=%p server=%p\n", static_cast<void*>(status), static_cast<void*>(sharedServer));
+            if (status != nullptr) {
+                const char* msg = nullptr;
+                OVMS_StatusDetails(status, &msg);
+                std::string details = (msg != nullptr) ? msg : "unknown error";
+                OVMS_StatusDelete(status);
+                throw std::runtime_error("OVMS_ServerNew failed in OVMSInferenceAdapter: " + details);
+            }
+        });
+        std::fprintf(stderr, "OVMSAdapter shared handle call_once exit server=%p\n", static_cast<void*>(sharedServer));
+        return sharedServer;
+    }
+
 public:
     // TODO Windows: Fix definition in header - does not compile in cpp.
     OVMSInferenceAdapter(const std::string& servableName, uint32_t servableVersion = 0, OVMS_Server* server = nullptr) :
@@ -62,7 +109,7 @@ class OVMSInferenceAdapter : public ::InferenceAdapter {
         if (nullptr != server) {
             this->cserver = server;
         } else {
-            OVMS_ServerNew(&this->cserver);
+            this->cserver = getSharedServerHandle();
         }
     }
     virtual ~OVMSInferenceAdapter();
diff --git a/mediapipe/calculators/ovms/openvinoinferencecalculator.cc b/mediapipe/calculators/ovms/openvinoinferencecalculator.cc
index c4c47f14700..98f880b0bf7 100644
--- a/mediapipe/calculators/ovms/openvinoinferencecalculator.cc
+++ b/mediapipe/calculators/ovms/openvinoinferencecalculator.cc
@@ -27,50 +27,33 @@
 #include "ovms.h"  // NOLINT
 #pragma GCC diagnostic push
 #pragma GCC diagnostic ignored "-Wdeprecated-declarations"
+#if !defined(OVMS_RUNTIME_DISABLE_TFLITE_TENSORS)
 #include "tensorflow/core/framework/tensor.h"
+#endif
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/port/canonical_errors.h"
 #include "mediapipe/framework/formats/tensor.h"
 #include "mediapipe/calculators/ovms/openvinoinferencecalculator.pb.h"
 #include "mediapipe/calculators/ovms/openvinoinferencecalculatoroptions.h"
 #include "mediapipe/calculators/ovms/openvinoinferenceutils.h"
+#if !defined(OVMS_RUNTIME_DISABLE_TFLITE_TENSORS)
 #include "tensorflow/lite/c/common.h"
+#endif
 #if (OVMS_DUMP_TO_FILE == 1)
 #include "openvinoinferencedumputils.h"
 #endif
 #pragma GCC diagnostic pop
+#if !defined(OVMS_RUNTIME_DISABLE_TFLITE_TENSORS)
 #pragma GCC diagnostic push
 #pragma GCC diagnostic ignored "-Wall"
 #include "tensorflow/lite/interpreter.h"
 #pragma GCC diagnostic pop
+#endif
 // here we need to decide if we have several calculators (1 for OVMS repository, 1-N inside mediapipe)
 // for the one inside OVMS repo it makes sense to reuse code from ovms lib
 namespace mediapipe {
 using std::endl;
 
-using TFSDataType = tensorflow::DataType;
-
-TFSDataType getPrecisionAsDataType(ov::element::Type_t precision) {
-    static std::unordered_map<ov::element::Type_t, TFSDataType> precisionMap{
-        {ov::element::Type_t::f32, TFSDataType::DT_FLOAT},
-        {ov::element::Type_t::f64, TFSDataType::DT_DOUBLE},
-        {ov::element::Type_t::f16, TFSDataType::DT_HALF},
-        {ov::element::Type_t::i64, TFSDataType::DT_INT64},
-        {ov::element::Type_t::i32, TFSDataType::DT_INT32},
-        {ov::element::Type_t::i16, TFSDataType::DT_INT16},
-        {ov::element::Type_t::i8, TFSDataType::DT_INT8},
-        {ov::element::Type_t::u64, TFSDataType::DT_UINT64},
-        {ov::element::Type_t::u16, TFSDataType::DT_UINT16},
-        {ov::element::Type_t::u8, TFSDataType::DT_UINT8},
-        {ov::element::Type_t::boolean, TFSDataType::DT_BOOL}
-    };
-    auto it = precisionMap.find(precision);
-    if (it == precisionMap.end()) {
-        return TFSDataType::DT_INVALID;
-    }
-    return it->second;
-}
-
 static Tensor::ElementType OVType2MPType(ov::element::Type_t precision) {
     static std::unordered_map<ov::element::Type_t, Tensor::ElementType> precisionMap{
 //        {ov::element::Type_t::f64, Tensor::ElementType::},
@@ -195,6 +178,30 @@ static Tensor convertOVTensor2MPTensor(const ov::Tensor& inputTensor) {
     return outputTensor;
 }
 
+#if !defined(OVMS_RUNTIME_DISABLE_TFLITE_TENSORS)
+using TFSDataType = tensorflow::DataType;
+
+TFSDataType getPrecisionAsDataType(ov::element::Type_t precision) {
+    static std::unordered_map<ov::element::Type_t, TFSDataType> precisionMap{
+        {ov::element::Type_t::f32, TFSDataType::DT_FLOAT},
+        {ov::element::Type_t::f64, TFSDataType::DT_DOUBLE},
+        {ov::element::Type_t::f16, TFSDataType::DT_HALF},
+        {ov::element::Type_t::i64, TFSDataType::DT_INT64},
+        {ov::element::Type_t::i32, TFSDataType::DT_INT32},
+        {ov::element::Type_t::i16, TFSDataType::DT_INT16},
+        {ov::element::Type_t::i8, TFSDataType::DT_INT8},
+        {ov::element::Type_t::u64, TFSDataType::DT_UINT64},
+        {ov::element::Type_t::u16, TFSDataType::DT_UINT16},
+        {ov::element::Type_t::u8, TFSDataType::DT_UINT8},
+        {ov::element::Type_t::boolean, TFSDataType::DT_BOOL}
+    };
+    auto it = precisionMap.find(precision);
+    if (it == precisionMap.end()) {
+        return TFSDataType::DT_INVALID;
+    }
+    return it->second;
+}
+
 ov::element::Type_t TFSPrecisionToIE2Precision(TFSDataType precision) {
     static std::unordered_map<TFSDataType, ov::element::Type_t> precisionMap{
         {TFSDataType::DT_DOUBLE, ov::element::Type_t::f64},
@@ -278,14 +285,17 @@ static ov::Tensor convertTFLiteTensor2OVTensor(const TfLiteTensor& t) {
     ov::Tensor result(datatype, shape, data);
     return result;
 }
+#endif
 
 class OpenVINOInferenceCalculator : public CalculatorBase {
     std::shared_ptr<::InferenceAdapter> session{nullptr};
     std::unordered_map<std::string, std::string> outputNameToTag;
     std::vector<std::string> input_order_list;
     std::vector<std::string> output_order_list;
+#if !defined(OVMS_RUNTIME_DISABLE_TFLITE_TENSORS)
     std::unique_ptr<tflite::Interpreter> interpreter_ = absl::make_unique<tflite::Interpreter>();
     bool initialized = false;
+#endif
 
 public:
     static absl::Status GetContract(CalculatorContract* cc) {
@@ -296,6 +306,12 @@ class OpenVINOInferenceCalculator : public CalculatorBase {
         RET_CHECK(ValidateCalculatorSettings(cc));
 
         for (const std::string& tag : cc->Inputs().GetTags()) {
+#if defined(OVMS_RUNTIME_DISABLE_TFLITE_TENSORS)
+            if (startsWith(tag, TFTENSORS_TAG) || startsWith(tag, TFTENSOR_TAG) || startsWith(tag, TFLITE_TENSORS_TAG) || startsWith(tag, TFLITE_TENSOR_TAG)) {
+                LOG(INFO) << "TensorFlow/TFLite packet tags are not supported in OVMS runtime-shared mode: " << tag;
+                RET_CHECK(false);
+            }
+#endif
             // could be replaced with absl::StartsWith when migrated to MP
             if (startsWith(tag, OVTENSORS_TAG)) {
                 LOG(INFO) << "setting input tag:" << tag << " to OVTensors";
@@ -309,6 +325,7 @@ class OpenVINOInferenceCalculator : public CalculatorBase {
             } else if (startsWith(tag, MPTENSOR_TAG)) {
                 LOG(INFO) << "setting input tag:" << tag << " to MPTensor";
                 cc->Inputs().Tag(tag).Set<Tensor>();
+#if !defined(OVMS_RUNTIME_DISABLE_TFLITE_TENSORS)
             } else if (startsWith(tag, TFTENSORS_TAG)) {
                 LOG(INFO) << "setting input tag:" << tag << " to TFTensors";
                 cc->Inputs().Tag(tag).Set<std::vector<tensorflow::Tensor>>();
@@ -321,12 +338,19 @@ class OpenVINOInferenceCalculator : public CalculatorBase {
             } else if (startsWith(tag, TFLITE_TENSOR_TAG)) {
                 LOG(INFO) << "setting input tag:" << tag << " to TFLITE_Tensor";
                 cc->Inputs().Tag(tag).Set<TfLiteTensor>();
+#endif
             } else {
                 LOG(INFO) << "setting input tag:" << tag << " to OVTensor";
                 cc->Inputs().Tag(tag).Set<ov::Tensor>();
             }
         }
         for (const std::string& tag : cc->Outputs().GetTags()) {
+#if defined(OVMS_RUNTIME_DISABLE_TFLITE_TENSORS)
+            if (startsWith(tag, TFTENSORS_TAG) || startsWith(tag, TFTENSOR_TAG) || startsWith(tag, TFLITE_TENSORS_TAG) || startsWith(tag, TFLITE_TENSOR_TAG)) {
+                LOG(INFO) << "TensorFlow/TFLite packet tags are not supported in OVMS runtime-shared mode: " << tag;
+                RET_CHECK(false);
+            }
+#endif
             if (startsWith(tag, OVTENSORS_TAG)) {
                 LOG(INFO) << "setting input tag:" << tag << " to std::vector<ov::Tensor>";
                 cc->Outputs().Tag(tag).Set<std::vector<ov::Tensor>>();
@@ -339,6 +363,7 @@ class OpenVINOInferenceCalculator : public CalculatorBase {
             } else if (startsWith(tag, MPTENSOR_TAG)) {
                 LOG(INFO) << "setting input tag:" << tag << " to MPTensor";
                 cc->Outputs().Tag(tag).Set<Tensor>();
+#if !defined(OVMS_RUNTIME_DISABLE_TFLITE_TENSORS)
             } else if (startsWith(tag, TFTENSORS_TAG)) {
                 LOG(INFO) << "setting input tag:" << tag << " to TFTensor";
                 cc->Outputs().Tag(tag).Set<std::vector<tensorflow::Tensor>>();
@@ -351,6 +376,7 @@ class OpenVINOInferenceCalculator : public CalculatorBase {
             } else if (startsWith(tag, TFLITE_TENSOR_TAG)) {
                 LOG(INFO) << "setting input tag:" << tag << " to TFLITE_Tensor";
                 cc->Outputs().Tag(tag).Set<TfLiteTensor>();
+#endif
             } else {
                 LOG(INFO) << "setting output tag:" << tag << " to OVTensor";
                 cc->Outputs().Tag(tag).Set<ov::Tensor>();
@@ -443,22 +469,28 @@ class OpenVINOInferenceCalculator : public CalculatorBase {
             try {
             if (startsWith(tag, OVTENSORS_TAG)) {
                 DESERIALIZE_TENSORS(ov::Tensor,);
+#if !defined(OVMS_RUNTIME_DISABLE_TFLITE_TENSORS)
             } else if (startsWith(tag, TFLITE_TENSORS_TAG)) {
                 DESERIALIZE_TENSORS(TfLiteTensor, convertTFLiteTensor2OVTensor);
+#endif
             } else if (startsWith(tag, MPTENSORS_TAG)) {
                 DESERIALIZE_TENSORS(Tensor, convertMPTensor2OVTensor);
             } else if (startsWith(tag, OVTENSOR_TAG)) {
                 auto& packet = cc->Inputs().Tag(tag).Get<ov::Tensor>();
                 input[realInputName] = packet;
+#if !defined(OVMS_RUNTIME_DISABLE_TFLITE_TENSORS)
             } else if (startsWith(tag, TFLITE_TENSOR_TAG)) {
                 auto& packet = cc->Inputs().Tag(tag).Get<TfLiteTensor>();
                 input[realInputName] = convertTFLiteTensor2OVTensor(packet);
+#endif
             } else if (startsWith(tag, MPTENSOR_TAG)) {
                 auto& packet = cc->Inputs().Tag(tag).Get<Tensor>();
                 input[realInputName] = convertMPTensor2OVTensor(packet);
+#if !defined(OVMS_RUNTIME_DISABLE_TFLITE_TENSORS)
             } else if (startsWith(tag, TFTENSOR_TAG)) {
                 auto& packet = cc->Inputs().Tag(tag).Get<tensorflow::Tensor>();
                 input[realInputName] = convertTFTensor2OVTensor(packet);
+#endif
             } else {
                 auto& packet = cc->Inputs().Tag(tag).Get<ov::Tensor>();
                 input[realInputName] = packet;
@@ -545,6 +577,7 @@ class OpenVINOInferenceCalculator : public CalculatorBase {
                 SERIALIZE_TENSORS(Tensor, convertOVTensor2MPTensor)
                 // no need to break since we only have one tag
                 // create concatenator calc
+#if !defined(OVMS_RUNTIME_DISABLE_TFLITE_TENSORS)
             } else if (startsWith(tag, TFLITE_TENSORS_TAG)) {
                 LOG(INFO) << "OVMS calculator will process vector<TfLiteTensor>";
                 auto outputStreamTensors = std::vector<TfLiteTensor>();
@@ -581,16 +614,19 @@ class OpenVINOInferenceCalculator : public CalculatorBase {
                 }
                 cc->Outputs().Tag(tag).AddPacket(MakePacket<std::vector<TfLiteTensor>>(std::move(outputStreamTensors)).At( cc->InputTimestamp()));
                 break;
+#endif
             } else if (startsWith(tag, OVTENSOR_TAG)) {
                 LOG(INFO) << "OVMS calculator will process ov::Tensor";
                 cc->Outputs().Tag(tag).Add(
                     new ov::Tensor(tensorIt->second),
                     cc->InputTimestamp());
+#if !defined(OVMS_RUNTIME_DISABLE_TFLITE_TENSORS)
             } else if (startsWith(tag, TFTENSOR_TAG)) {
                 LOG(INFO) << "OVMS calculator will process tensorflow::Tensor";
                 cc->Outputs().Tag(tag).Add(
                     new tensorflow::Tensor(convertOVTensor2TFTensor(tensorIt->second)),
                     cc->InputTimestamp());
+#endif
             } else if (startsWith(tag, MPTENSOR_TAG)) {
                 LOG(INFO) << "OVMS calculator will process mediapipe::Tensor";
                 cc->Outputs().Tag(tag).Add(
diff --git a/mediapipe/calculators/ovms/openvinoinferencecalculator.h b/mediapipe/calculators/ovms/openvinoinferencecalculator.h
index 8a92b8120c4..9c9e5cc7f86 100644
--- a/mediapipe/calculators/ovms/openvinoinferencecalculator.h
+++ b/mediapipe/calculators/ovms/openvinoinferencecalculator.h
@@ -22,10 +22,12 @@
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/port/canonical_errors.h"
 #pragma GCC diagnostic pop
+#if !defined(OVMS_RUNTIME_DISABLE_TFLITE_TENSORS)
 #pragma GCC diagnostic push
 #pragma GCC diagnostic ignored "-Wall"
 #include "tensorflow/lite/interpreter.h"
 #pragma GCC diagnostic pop
+#endif
 class InferenceAdapter;
 namespace mediapipe {
 class OpenVINOInferenceCalculator : public CalculatorBase {
@@ -33,8 +35,10 @@ class OpenVINOInferenceCalculator : public CalculatorBase {
     std::unordered_map<std::string, std::string> outputNameToTag;
     std::vector<std::string> input_order_list;
     std::vector<std::string> output_order_list;
+#if !defined(OVMS_RUNTIME_DISABLE_TFLITE_TENSORS)
     std::unique_ptr<tflite::Interpreter> interpreter_ = absl::make_unique<tflite::Interpreter>();
     bool initialized = false;
+#endif
 public:
     static absl::Status GetContract(CalculatorContract* cc);
     absl::Status Close(CalculatorContext* cc) override final;
diff --git a/mediapipe/calculators/ovms/openvinomodelserversessioncalculator.cc b/mediapipe/calculators/ovms/openvinomodelserversessioncalculator.cc
index e6924801a06..c3da05efebf 100644
--- a/mediapipe/calculators/ovms/openvinomodelserversessioncalculator.cc
+++ b/mediapipe/calculators/ovms/openvinomodelserversessioncalculator.cc
@@ -22,6 +22,12 @@
 #include <sstream>
 #include <unordered_map>
 
+#ifdef _WIN32
+#include <Windows.h>
+#else
+#include <dlfcn.h>
+#endif
+
 #include <openvino/openvino.hpp>
 
 #include "ovms.h"           // NOLINT
@@ -36,6 +42,25 @@
 namespace mediapipe {
 using ovms::OVMSInferenceAdapter;
 
+namespace {
+OVMS_Server* getInheritedServerHandleFromRuntime() {
+    using GetterFn = void* (*)();
+#ifdef _WIN32
+    HMODULE runtimeHandle = GetModuleHandleA("ovms_mediapipe_runtime_shared.dll");
+    if (runtimeHandle == nullptr) {
+        return nullptr;
+    }
+    auto* fn = reinterpret_cast<GetterFn>(GetProcAddress(runtimeHandle, "OVMS_MPGetExternalServerHandle"));
+#else
+    auto* fn = reinterpret_cast<GetterFn>(dlsym(RTLD_DEFAULT, "OVMS_MPGetExternalServerHandle"));
+#endif
+    if (fn == nullptr) {
+        return nullptr;
+    }
+    return static_cast<OVMS_Server*>(fn());
+}
+}  // namespace
+
 const std::string SESSION_TAG{"SESSION"};
 ov::Core UNUSED_OV_CORE;
 
@@ -133,7 +158,15 @@ absl::Status OpenVINOModelServerSessionCalculator::Close(CalculatorContext* cc)
     LOG(INFO) << "OpenVINOModelServerSessionCalculator Close";
     return absl::OkStatus();
 }
+// Open() prepares an inference adapter and publishes it as a side packet.
+// Flow:
+// 1) If server_config is provided, it lazily creates/starts OVMS once (guarded by loadingMtx)
+//    and validates that the server is live/ready.
+// 2) If server_config is empty, it reuses a server handle inherited from runtime glue.
+// 3) It builds OVMSInferenceAdapter for (servable_name, servable_version), loads model metadata,
+//    and stores the adapter in SESSION_TAG output side packet.
 absl::Status OpenVINOModelServerSessionCalculator::Open(CalculatorContext* cc) {
+    cc->SetOffset(TimestampDiff(0));
     LOG(INFO) << "OpenVINOModelServerSessionCalculator Open start";
     for (CollectionItemId id = cc->Inputs().BeginId();
          id < cc->Inputs().EndId(); ++id) {
@@ -147,19 +180,20 @@ absl::Status OpenVINOModelServerSessionCalculator::Open(CalculatorContext* cc) {
             cc->OutputSidePackets().Get(id).Set(cc->InputSidePackets().Get(id));
         }
     }
-    cc->SetOffset(TimestampDiff(0));
 
     const auto& options = cc->Options<OpenVINOModelServerSessionCalculatorOptions>();
     // if config is in calc then we start the server
     LOG(INFO) << "Will check if we want to start server";
+    bool isServerLive = false;
     if (!options.server_config().empty()) {
-        // Lock access to server from multiple calculator instances during the model loading phase
+        // Lock access to shared server handle from multiple calculator instances.
         std::unique_lock<std::mutex> lk(OpenVINOModelServerSessionCalculator::loadingMtx);
-        bool isServerReady = false;
-        bool isServerLive = false;
-        OVMS_ServerNew(&cserver);
-
+        if (cserver == nullptr) {
+            ASSERT_CAPI_STATUS_NULL(OVMS_ServerNew(&cserver));
+        }
         ASSERT_CAPI_STATUS_NULL(OVMS_ServerLive(cserver, &isServerLive));
+        // if config is in calc then we start the server
+        bool isServerReady = false;
 
         if (triedToStartOVMS) {
             RET_CHECK(isServerLive);
@@ -180,13 +214,17 @@ absl::Status OpenVINOModelServerSessionCalculator::Open(CalculatorContext* cc) {
             RET_CHECK(isServerReady);
             LOG(INFO) << "Server started";
         }
+    } else {
+        if (cserver == nullptr) {
+            cserver = getInheritedServerHandleFromRuntime();
+        }
     }
     const std::string& servableName = options.servable_name();
     const std::string& servableVersionStr = options.servable_version();
     auto servableVersionOpt = stou32(servableVersionStr);
     // 0 means default
     uint32_t servableVersion = servableVersionOpt.value_or(0);
-    auto session = std::make_shared<OVMSInferenceAdapter>(servableName, servableVersion);
+    auto session = std::make_shared<OVMSInferenceAdapter>(servableName, servableVersion, cserver);
     try {
         session->loadModel(nullptr, UNUSED_OV_CORE, "UNUSED", {});
     } catch (const std::exception& e) {
diff --git a/third_party/BUILD b/third_party/BUILD
index 783762f4937..4ac98c2ddd1 100644
--- a/third_party/BUILD
+++ b/third_party/BUILD
@@ -51,10 +51,10 @@ cc_library(
             "@com_github_glog_glog//:glog",
         ],
         "//mediapipe:windows": [
-            "@com_github_glog_glog_windows//:glog",
+            "@com_github_glog_glog//:glog",
         ],
         "//conditions:default": [
-            "@com_github_glog_glog//:glog",
+            "@com_github_glog_glog_no_gflags//:glog",
         ],
     }),
 )
diff --git a/third_party/model_api/model_api.bzl b/third_party/model_api/model_api.bzl
index cd4b2c2565d..ab5045fac7e 100644
--- a/third_party/model_api/model_api.bzl
+++ b/third_party/model_api/model_api.bzl
@@ -51,7 +51,7 @@ cmake(
     tags = ["requires-network"],
     deps = [
         "@mediapipe//mediapipe/framework/port:opencv_core",
-        "@mediapipe//third_party:openvino",
+        "@windows_openvino//:openvino_new_headers",
     ],
 )
 
@@ -59,7 +59,7 @@ cc_library(
     name = "model_api",
     deps = [
         "@mediapipe//mediapipe/framework/port:opencv_core",
-        "@mediapipe//third_party:openvino",
+        "@windows_openvino//:openvino_new_headers",
         ":model_api_cmake",
     ],
     visibility = ["//visibility:public"],