diff --git a/src/sst/core/config.cc b/src/sst/core/config.cc
index 63939421f..524cf6daa 100644
--- a/src/sst/core/config.cc
+++ b/src/sst/core/config.cc
@@ -798,7 +798,11 @@ static const struct sstLongOpts_s sstOptions[] = {
         "verbose", 'v', "level", "Verbosity level to determine what information about core runtime is printed",
         &ConfigHelper::incrVerbose, &ConfigHelper::setVerbosity, true),
     DEF_ARG(
-        "num_threads", 'n', "NUM", "Number of parallel threads to use per rank", &ConfigHelper::setNumThreads, true),
+        "num-threads", 'n', "NUM", "Number of parallel threads to use per rank", &ConfigHelper::setNumThreads, true),
+    DEF_ARG(
+        "num_threads", 0, "NUM",
+        "[Deprecated] Number of parallel threads to use per rank (deprecated, please use --num-threads or -n instead)",
+        &ConfigHelper::setNumThreads, true),
     DEF_ARG(
         "sdl-file", 0, "FILE",
         "Specify SST Configuration file.  Note: this is most often done by just specifying the file without an option.",
diff --git a/src/sst/core/main.cc b/src/sst/core/main.cc
index 0c595aba3..fcf634294 100644
--- a/src/sst/core/main.cc
+++ b/src/sst/core/main.cc
@@ -640,6 +640,20 @@ main(int argc, char* argv[])
     }
 
     ////// End ConfigGraph Creation //////
+#ifdef SST_CONFIG_HAVE_MPI
+    // If we did a parallel load, check to make sure that all the
+    // ranks have the same thread count set (the python can change the
+    // thread count if not specified on the command line
+    if ( cfg.parallel_load() ) {
+        uint32_t max_thread_count = 0;
+        uint32_t my_thread_count  = cfg.num_threads();
+        MPI_Allreduce(&my_thread_count, &max_thread_count, 1, MPI_UINT32_T, MPI_MAX, MPI_COMM_WORLD);
+        if ( my_thread_count != max_thread_count ) {
+            g_output.fatal(
+                CALL_INFO, 1, "Thread counts do no match across ranks for configuration using parallel loading\n");
+        }
+    }
+#endif
 
     ////// Start Partitioning //////
     double start_part = sst_get_cpu_time();
diff --git a/src/sst/core/model/python/pymodel.cc b/src/sst/core/model/python/pymodel.cc
index f4ae15dc9..e951c6b2f 100644
--- a/src/sst/core/model/python/pymodel.cc
+++ b/src/sst/core/model/python/pymodel.cc
@@ -397,7 +397,7 @@ setSSTThreadCount(PyObject* UNUSED(self), PyObject* args)
     Config* cfg     = gModel->getConfig();
     long    oldNThr = cfg->num_threads();
     long    nThr    = SST_ConvertToCppLong(args);
-    if ( nThr > 0 && nThr <= oldNThr ) gModel->setConfigEntryFromModel("num_threads", std::to_string(nThr));
+    gModel->setConfigEntryFromModel("num_threads", std::to_string(nThr));
     return SST_ConvertToPythonLong(oldNThr);
 }