diff --git a/mlx_lm/tokenizer_utils.py b/mlx_lm/tokenizer_utils.py
index c7e50fbe7..a10d7cb55 100644
--- a/mlx_lm/tokenizer_utils.py
+++ b/mlx_lm/tokenizer_utils.py
@@ -358,7 +358,7 @@ def add_eos_token(self, token: str):
         self._eos_token_ids.add(token_id)
 
     def _find(self, tokens, sequence, start=None, end=None, reverse=False):
-        start = start or 0
+        start = max(start or 0, 0)
         end = end or len(tokens)
         outer_loop = (
             range(end - len(sequence), start - 1, -1)
diff --git a/tests/test_tokenizers.py b/tests/test_tokenizers.py
index 54906af1c..cce4dd6a3 100644
--- a/tests/test_tokenizers.py
+++ b/tests/test_tokenizers.py
@@ -109,6 +109,18 @@ def test_thinking(self):
         self.assertIsNone(tokenizer.think_start_id)
         self.assertIsNone(tokenizer.think_end_id)
 
+    def test_think_search_short_prompt_negative_start(self):
+        # Regression for #1326: the server computes start = len(prompt) - 11 and passes
+        # it to rfind_think_start. For prompts shorter than 11 tokens that start is
+        # negative; the think-token search must clamp it instead of indexing past the
+        # token list (IndexError, which the server surfaces as a 404).
+        tokenizer = load_tokenizer("mlx-community/Qwen3-4B-4bit")
+        self.assertTrue(tokenizer.has_thinking)
+        short_tokens = [0, 1, 2]  # 3-token prompt with no think-start present
+        start = len(short_tokens) - 11  # = -8, exactly what server.py passes
+        self.assertEqual(tokenizer.rfind_think_start(short_tokens, start=start), -1)
+        self.assertEqual(tokenizer.find_think_start(short_tokens, start=start), -1)
+
 
 if __name__ == "__main__":
     unittest.main()