diff --git a/pkg/llms_from_scratch/qwen3.py b/pkg/llms_from_scratch/qwen3.py
index 3e9f726..11542a6 100644
--- a/pkg/llms_from_scratch/qwen3.py
+++ b/pkg/llms_from_scratch/qwen3.py
@@ -536,14 +536,14 @@ class Qwen3Tokenizer:
         self._special_to_id = {t: self._tok.token_to_id(t) for t in self._SPECIALS}
 
         self.pad_token_id = self._special_to_id.get("<|endoftext|>")
-        self.eos_token_id = self.pad_token_id
 
-        if repo_id and "Base" not in repo_id:
-            eos_token = "<|im_end|>"
+        # Match HF behavior: chat model → <|im_end|>, base model → <|endoftext|>
+        fname = tok_file.name.lower()
+        if "base" in fname and "reasoning" not in fname:
+            self.eos_token = "<|endoftext|>"
         else:
-            eos_token = "<|endoftext|>"
-        if eos_token in self._special_to_id:
-            self.eos_token_id = self._special_to_id[eos_token]
+            self.eos_token = "<|im_end|>"
+        self.eos_token_id = self._special_to_id.get(self.eos_token)
 
     def encode(self, text, chat_wrapped=None):
         if chat_wrapped is None: