diff --git a/engine.py b/engine.py
index 766ed87..a66d95f 100644
--- a/engine.py
+++ b/engine.py
@@ -50,6 +50,19 @@ def load_model() -> bool:
             _is_turbo = False
 
         _sample_rate = 24000
+
+        # Convert weights to fp16. Done once at load time so the warmup
+        # covers the right dtypes and there's no per-call casting overhead.
+        if torch.cuda.is_available():
+            try:
+                for attr in ("t3", "s3gen", "ve"):
+                    m = getattr(chatterbox_model, attr, None)
+                    if m is not None:
+                        m.half()
+                logger.info("Model converted to fp16")
+            except Exception:
+                logger.warning("fp16 conversion failed, running in fp32", exc_info=True)
+
         logger.info("Model loaded successfully")
         return True
     except Exception: