diff --git a/DESCRIPTION b/DESCRIPTION
index a5819855..af5c6967 100644
--- a/DESCRIPTION
+++ b/DESCRIPTION
@@ -53,6 +53,7 @@ Suggests:
     mgcv,
     mirai (>= 2.4.0),
     modeldata,
+    mori,
     probably,
     scales,
     spelling,
diff --git a/R/loop_over_all_stages.R b/R/loop_over_all_stages.R
index 4cf35eb4..44b6acbc 100644
--- a/R/loop_over_all_stages.R
+++ b/R/loop_over_all_stages.R
@@ -26,6 +26,9 @@
   }
 
   split <- resamples$splits[[1]]
+  if (!is.null(static$mori_shared_name)) {
+    split$data <- mori::map_shared(static$mori_shared_name)
+  }
   split_labs <- resamples |>
     dplyr::select(dplyr::starts_with("id"))
   # Determine if we're using prediction-based or model-based metrics
diff --git a/R/tune_grid_loop.R b/R/tune_grid_loop.R
index 1e146bd1..adf7c6dc 100644
--- a/R/tune_grid_loop.R
+++ b/R/tune_grid_loop.R
@@ -108,6 +108,20 @@ tune_grid_loop <- function(
     inds <- vec_list_rowwise(inds)
   }
 
+  # Share the full training data via mori to avoid serializing it to every
+  # worker. Instead, workers receive lightweight split objects (indices + schema
+  # only) and map the shared memory region directly.
+  mori_ref <- NULL
+  if (strategy != "sequential" && rlang::is_installed("mori")) {
+    full_data <- resamples[[1]]$splits[[1]]$data
+    mori_ref <- mori::share(full_data)
+    static$mori_shared_name <- mori::shared_name(mori_ref)
+    resamples <- purrr::map(resamples, function(res_row) {
+      res_row$splits[[1]]$data <- full_data[0L, ]
+      res_row
+    })
+  }
+
   cl <- loop_call(control$parallel_over, strategy, par_opt)
   res <- rlang::eval_bare(cl)