diff --git a/src/export_cache.cpp b/src/export_cache.cpp
index c13f14a..a9f3afb 100644
--- a/src/export_cache.cpp
+++ b/src/export_cache.cpp
@@ -107,6 +107,7 @@ cache_t* pypluginCache_init(
     py::function cache_init_hook, py::function cache_hit_hook,
     py::function cache_miss_hook, py::function cache_eviction_hook,
     py::function cache_remove_hook, py::function cache_free_hook) {
+  py::gil_scoped_acquire acquire;
   // Initialize base cache structure with exception safety
   cache_t* cache = nullptr;
   std::unique_ptr<pypluginCache_params_t, PypluginCacheParamsDeleter> params;
@@ -163,20 +164,28 @@ cache_t* pypluginCache_init(
 }
 
 static void pypluginCache_free(cache_t* cache) {
-  if (!cache || !cache->eviction_params) {
+  if (!cache) {
+    return;
+  }
+  py::gil_scoped_acquire acquire;
+  if (!cache->eviction_params) {
+    // No params, just free the cache structure
+    cache_struct_free(cache);
     return;
   }
 
-  // Use smart pointer for automatic cleanup
-  std::unique_ptr<pypluginCache_params_t, PypluginCacheParamsDeleter> params(
-      static_cast<pypluginCache_params_t*>(cache->eviction_params));
-
+  auto* raw_params = static_cast<pypluginCache_params_t*>(cache->eviction_params);
+  cache->eviction_params = nullptr;
   // The smart pointer destructor will handle cleanup automatically
+  std::unique_ptr<pypluginCache_params_t, PypluginCacheParamsDeleter> params(raw_params);
+  params.reset();
+
   cache_struct_free(cache);
 }
 
 static bool pypluginCache_get(cache_t* cache, const request_t* req) {
   bool hit = cache_get_base(cache, req);
+  py::gil_scoped_acquire acquire;
   pypluginCache_params_t* params =
       (pypluginCache_params_t*)cache->eviction_params;
 
@@ -204,6 +213,7 @@ static cache_obj_t* pypluginCache_to_evict(cache_t* cache,
 }
 
 static void pypluginCache_evict(cache_t* cache, const request_t* req) {
+  py::gil_scoped_acquire acquire;
   pypluginCache_params_t* params =
       (pypluginCache_params_t*)cache->eviction_params;
 
@@ -223,6 +233,7 @@ static void pypluginCache_evict(cache_t* cache, const request_t* req) {
 }
 
 static bool pypluginCache_remove(cache_t* cache, const obj_id_t obj_id) {
+  py::gil_scoped_acquire acquire;
   pypluginCache_params_t* params =
       (pypluginCache_params_t*)cache->eviction_params;
 
@@ -568,7 +579,8 @@ void export_cache(py::module& m) {
             bytes_req > 0 ? 1.0 - (double)bytes_hit / bytes_req : 0.0;
         return std::make_tuple(obj_miss_ratio, byte_miss_ratio);
       },
-      "cache"_a, "reader"_a, "start_req"_a = 0, "max_req"_a = -1);
+      "cache"_a, "reader"_a, "start_req"_a = 0, "max_req"_a = -1,
+      py::call_guard<py::gil_scoped_release>());
 }
 
 }  // namespace libcachesim
diff --git a/tests/test_gil.py b/tests/test_gil.py
new file mode 100644
index 0000000..4a9d9eb
--- /dev/null
+++ b/tests/test_gil.py
@@ -0,0 +1,50 @@
+import pytest
+import libcachesim as lcs
+import threading
+import time
+
+S3_URI = "s3://cache-datasets/cache_dataset_oracleGeneral/2007_msr/msr_hm_0.oracleGeneral.zst"
+
+def run_heavy_simulation(name):
+    # Create a large synthetic trace
+    reader = lcs.TraceReader(trace=S3_URI)
+    cache = lcs.LRU(cache_size=1024*1024)
+
+    print(f"Thread {name} starting simulation...")
+    start = time.time()
+    # Call C++ core logic
+    lcs.Util.process_trace(cache, reader)
+    end = time.time()
+    print(f"Thread {name} completed in {end - start:.2f}s")
+
+def test_gil_release():
+    """
+    Test to verify that the GIL is released during heavy C++ processing.
+    We run two threads that perform heavy simulations and measure total time.
+    If the total time is close to the single-thread time, it indicates GIL release.
+    If the total time is close to double the single-thread time, it indicates GIL is still held.
+    """
+    # --- Experiment start ---
+
+    # test single-thread time for reference
+    start_single = time.time()
+    run_heavy_simulation("Single")
+    end_single = time.time()
+    single_thread_time = end_single - start_single
+    print(f"\nSingle-thread time: {single_thread_time:.2f}s")
+
+    start_total = time.time()
+
+    t1 = threading.Thread(target=run_heavy_simulation, args=("A",))
+    t2 = threading.Thread(target=run_heavy_simulation, args=("B",))
+
+    t1.start()
+    t2.start()
+
+    t1.join()
+    t2.join()
+
+    end_total = time.time()
+    print(f"\nTotal elapsed time: {end_total - start_total:.2f}s")
+
+    assert single_thread_time * 1.5 > (end_total - start_total), "GIL release test failed: Total time should be close to single-thread time if GIL is released."