diff --git a/src/h3_blackwell_runtime/conditioning.py b/src/h3_blackwell_runtime/conditioning.py
index 8103875..cc7cb05 100644
--- a/src/h3_blackwell_runtime/conditioning.py
+++ b/src/h3_blackwell_runtime/conditioning.py
@@ -9,13 +9,9 @@ class H3PromptTokenizer:
     """Tokenize raw H3 prompt text without Qwen chat-template tokens."""
 
     def __init__(self, tokenizer_dir: str | Path):
-        from transformers import Qwen2Tokenizer
+        from transformers import AutoTokenizer
 
-        tokenizer_dir = Path(tokenizer_dir)
-        self.tokenizer = Qwen2Tokenizer(
-            vocab_file=str(tokenizer_dir / "vocab.json"),
-            merges_file=str(tokenizer_dir / "merges.txt"),
-        )
+        self.tokenizer = AutoTokenizer.from_pretrained(str(Path(tokenizer_dir)), local_files_only=True)
 
     def __call__(self, prompt: str, *, device: torch.device | str = "cuda") -> torch.Tensor:
         if not prompt:
diff --git a/src/h3_blackwell_runtime/qwen25_tokenizer/config.json b/src/h3_blackwell_runtime/qwen25_tokenizer/config.json
new file mode 100644
index 0000000..8863755
--- /dev/null
+++ b/src/h3_blackwell_runtime/qwen25_tokenizer/config.json
@@ -0,0 +1,5 @@
+{
+  "model_type": "qwen2",
+  "tokenizer_class": "Qwen2Tokenizer",
+  "vocab_size": 151936
+}