diff --git a/kt-kernel/python/utils/llamafile.py b/kt-kernel/python/utils/llamafile.py
index f39704bc..68dce644 100644
--- a/kt-kernel/python/utils/llamafile.py
+++ b/kt-kernel/python/utils/llamafile.py
@@ -217,3 +217,6 @@ def load_weights(self, physical_to_logical_map_cpu: Optional[torch.Tensor] = Non
         # Load weights
         self.cpu_infer.submit(self.moe.load_weights_task(physical_to_logical_map_cpu.data_ptr()))
         self.cpu_infer.sync()
+
+        # Drop original weights after loading
+        self.weights_to_keep = None
\ No newline at end of file