diff --git a/translate_epub.py b/translate_epub.py
index e948905..1e3dd5f 100644
--- a/translate_epub.py
+++ b/translate_epub.py
@@ -205,7 +205,7 @@ def load_model(model_id: str, load_in_8bit: bool = True) -> None:
     global _model, _tokenizer
 
     import torch
-    from transformers import AutoModelForCausalLM, AutoTokenizer
+    from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
 
     print(f"\n🔄 Loading model: {model_id}")
     print(f"   Quantization : {'INT8 (bitsandbytes)' if load_in_8bit else 'auto (float16/bfloat16)'}")
@@ -221,10 +221,11 @@ def load_model(model_id: str, load_in_8bit: bool = True) -> None:
 
     if load_in_8bit:
         # INT8 quantization via bitsandbytes — fits ~72 GB for a 72B model
+        quantization_config = BitsAndBytesConfig(load_in_8bit=True)
         _model = AutoModelForCausalLM.from_pretrained(
             model_id,
             device_map="auto",
-            load_in_8bit=True,
+            quantization_config=quantization_config,
             trust_remote_code=True,
             torch_dtype=torch.float16,
         )