diff --git a/src/h3_blackwell_runtime/conditioning.py b/src/h3_blackwell_runtime/conditioning.py
index b4d5f79..8103875 100644
--- a/src/h3_blackwell_runtime/conditioning.py
+++ b/src/h3_blackwell_runtime/conditioning.py
@@ -9,9 +9,13 @@ class H3PromptTokenizer:
     """Tokenize raw H3 prompt text without Qwen chat-template tokens."""
 
     def __init__(self, tokenizer_dir: str | Path):
-        from transformers import AutoTokenizer
+        from transformers import Qwen2Tokenizer
 
-        self.tokenizer = AutoTokenizer.from_pretrained(str(tokenizer_dir), local_files_only=True)
+        tokenizer_dir = Path(tokenizer_dir)
+        self.tokenizer = Qwen2Tokenizer(
+            vocab_file=str(tokenizer_dir / "vocab.json"),
+            merges_file=str(tokenizer_dir / "merges.txt"),
+        )
 
     def __call__(self, prompt: str, *, device: torch.device | str = "cuda") -> torch.Tensor:
         if not prompt: