diff --git a/src/main.rs b/src/main.rs
index 4ea02d1..cd19753 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -254,6 +254,17 @@ fn run() -> Result<()> {
         path: last_model_path.clone(),
     };
 
+    // Also ensure cleanup on panic: install a panic hook that removes .last_model, then chains
+    {
+        let last_for_panic = last_model_path.clone();
+        let prev_hook = std::panic::take_hook();
+        std::panic::set_hook(Box::new(move |info| {
+            let _ = std::fs::remove_file(&last_for_panic);
+            // chain to default/previous hook for normal panic reporting
+            prev_hook(info);
+        }));
+    }
+
     // Select backend
     let requested = match args.gpu_backend {
         GpuBackendCli::Auto => BackendKind::Auto,
@@ -740,6 +751,8 @@ fn run() -> Result<()> {
         }
     }
 
+    // Final best-effort cleanup of .last_model on normal exit
+    let _ = std::fs::remove_file(&last_model_path);
     Ok(())
 }