diff --git a/tests/test_gemma_added_tokens.py b/tests/test_gemma_added_tokens.py
new file mode 100644
index 000000000..e6bb3d22c
--- /dev/null
+++ b/tests/test_gemma_added_tokens.py
@@ -0,0 +1,29 @@
+import ast
+import json
+from pathlib import Path
+from types import SimpleNamespace
+import tempfile
+import unittest
+from unittest.mock import Mock
+
+source = Path(__file__).resolve().parents[1] / "utils/convert-bitnet-embedding-to-gguf.py"
+tree = ast.parse(source.read_text())
+methods = [node for node in tree.body if isinstance(node, ast.FunctionDef) and node.name in {"_set_vocab_gemma3", "_does_token_look_special"}]
+module = ast.Module(body=[ast.ImportFrom(module="__future__", names=[ast.alias(name="annotations")], level=0)] + methods, type_ignores=[])
+ast.fix_missing_locations(module)
+ns = {"json": json, "gguf": SimpleNamespace(TokenType=SimpleNamespace(UNUSED=5, CONTROL=3, USER_DEFINED=4, BYTE=6, UNKNOWN=2, NORMAL=1), SpecialVocab=lambda *a, **k: SimpleNamespace(add_to_gguf=lambda writer: None))}
+exec(compile(module, str(source), "exec"), ns)
+
+class GemmaAddedTokensTests(unittest.TestCase):
+    def test_added_tokens_outside_base_vocab_keep_ids_and_content(self):
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            (root / "tokenizer.json").write_text(json.dumps({"model": {"vocab": {"<unk>": 0, "base": 1}, "merges": []}}))
+            tokenizer = SimpleNamespace(added_tokens_decoder={2: SimpleNamespace(content="<eos>", special=True), 3: SimpleNamespace(content="custom", special=False)})
+            writer = Mock()
+            ns["_set_vocab_gemma3"](writer, root, tokenizer, 5)
+            writer.add_token_list.assert_called_once_with([b"<unk>", b"base", b"<eos>", b"custom", b"[PAD4]"])
+            writer.add_token_types.assert_called_once_with([2, 1, 3, 4, 5])
+            writer.add_token_scores.assert_called_once_with([0.0, 0.0, -1000.0, -1000.0, -10000.0])
+
+if __name__ == "__main__": unittest.main()
diff --git a/utils/convert-bitnet-embedding-to-gguf.py b/utils/convert-bitnet-embedding-to-gguf.py
index 3a9e48924..d7b7e1786 100644
--- a/utils/convert-bitnet-embedding-to-gguf.py
+++ b/utils/convert-bitnet-embedding-to-gguf.py
@@ -284,6 +284,9 @@ def _set_vocab_gemma3(gguf_writer: gguf.GGUFWriter, dir_model: Path,
     # Build token arrays
     reverse_vocab = {v: k for k, v in bpe_vocab.items()}
     added_tokens_decoder = tokenizer.added_tokens_decoder
+    # Added tokens can live outside tokenizer.json's base BPE vocabulary.
+    # Preserve their actual IDs/content before filling genuinely unused slots.
+    reverse_vocab.update({id_: token.content for id_, token in added_tokens_decoder.items()})
 
     tokens: list[bytes] = []
     scores: list[float] = []