diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index a8ddbfed2a..e1c13315f7 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -146,6 +146,7 @@ class FastLanguageModel(FastLlamaModel):
         disable_log_stats = True,
         qat_scheme = None,
         load_in_fp8 = False,  # fp8 LoRA (True, False, 'block')
+        unsloth_tiled_mlp = False,
         *args,
         **kwargs,
     ):
@@ -512,6 +513,7 @@ class FastLanguageModel(FastLlamaModel):
                 disable_log_stats = disable_log_stats,
                 qat_scheme = qat_scheme,
                 load_in_fp8 = load_in_fp8,
+                unsloth_tiled_mlp = unsloth_tiled_mlp,
                 *args,
                 **kwargs,
             )
@@ -609,8 +611,10 @@ class FastLanguageModel(FastLlamaModel):
 
         # Patch Tiled MLP
         # to turn on set UNSLOTH_TILED_MLP to "arctic", "target", or "target:{GB}""
-        patch_tiled_mlp_choice = os.environ.get("UNSLOTH_TILED_MLP", "0")
-        if patch_tiled_mlp_choice != "0":
+        patch_tiled_mlp_choice = os.environ.get(
+            "UNSLOTH_TILED_MLP", "arctic" if unsloth_tiled_mlp else "0"
+        )
+        if patch_tiled_mlp_choice != "0" or unsloth_tiled_mlp:
             patch_tiled_mlp(model, patch_options_str = patch_tiled_mlp_choice)
 
         return model, tokenizer
@@ -674,6 +678,7 @@ class FastModel(FastBaseModel):
         disable_log_stats = True,
         qat_scheme = None,
         load_in_fp8 = False,  # fp8 LoRA (True, False, 'block')
+        unsloth_tiled_mlp = False,
         *args,
         **kwargs,
     ):
@@ -1240,8 +1245,10 @@ class FastModel(FastBaseModel):
 
         # Patch Tiled MLP
         # to turn on set UNSLOTH_TILED_MLP to "arctic", "target", or "target:{GB}""
-        patch_tiled_mlp_choice = os.environ.get("UNSLOTH_TILED_MLP", "0")
-        if patch_tiled_mlp_choice != "0":
+        patch_tiled_mlp_choice = os.environ.get(
+            "UNSLOTH_TILED_MLP", "arctic" if unsloth_tiled_mlp else "0"
+        )
+        if patch_tiled_mlp_choice != "0" or unsloth_tiled_mlp:
             patch_tiled_mlp(model, patch_options_str = patch_tiled_mlp_choice)
 
         return model, tokenizer