diff --git a/tests/test_bugfix_generate_flush_optional.py b/tests/test_bugfix_generate_flush_optional.py
new file mode 100644
index 0000000..6ab3736
--- /dev/null
+++ b/tests/test_bugfix_generate_flush_optional.py
@@ -0,0 +1,85 @@
+import importlib
+import importlib.util
+import os
+import unittest
+from unittest import mock
+
+
+class GenerateFlushBugfixOptionalTests(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        enabled = os.getenv("VOXMLX_ENABLE_MLX_RUNTIME_TESTS", "").strip().lower() in {
+            "1",
+            "true",
+            "yes",
+            "on",
+        }
+        if not enabled:
+            raise unittest.SkipTest(
+                "Set VOXMLX_ENABLE_MLX_RUNTIME_TESTS=1 to run MLX runtime optional tests"
+            )
+        if importlib.util.find_spec("mlx") is None:
+            raise unittest.SkipTest("mlx is required for runtime optional tests")
+
+    def test_generate_flushes_last_pending_non_eos_token(self):
+        import mlx.core as mx
+        import numpy as np
+
+        generate_module = importlib.import_module("voxmlx.generate")
+
+        class _FakeLanguageModel:
+            def __init__(self, hidden_dim: int, vocab_size: int):
+                self.layers = [object()]
+                self._hidden_dim = hidden_dim
+                self._vocab_size = vocab_size
+
+            def embed(self, input_ids):
+                bsz, seq = input_ids.shape
+                return mx.zeros((bsz, seq, self._hidden_dim), dtype=mx.float32)
+
+        class _FakeModel:
+            def __init__(self):
+                self.hidden_dim = 8
+                self.vocab_size = 128
+                self.n_audio = 4  # prefix=2, loop runs 2 steps
+                self._decode_tokens = [11, 12, 13]  # prefill + 2 steps
+                self._decode_calls = 0
+                self.language_model = _FakeLanguageModel(self.hidden_dim, self.vocab_size)
+
+            def time_embedding(self, t):
+                return mx.zeros((1, self.hidden_dim), dtype=mx.float32)
+
+            def encode(self, mel):
+                del mel
+                return mx.zeros((self.n_audio, self.hidden_dim), dtype=mx.float32)
+
+            def decode(self, embeddings, t_cond, mask=None, cache=None):
+                del t_cond, mask, cache
+                token = self._decode_tokens[self._decode_calls]
+                self._decode_calls += 1
+                seq = embeddings.shape[1]
+                logits = np.full((1, seq, self.vocab_size), -1000.0, dtype=np.float32)
+                logits[:, :, token] = 1000.0
+                return mx.array(logits)
+
+        model = _FakeModel()
+        with mock.patch.object(generate_module, "load_audio", return_value=np.zeros(1, dtype=np.float32)), mock.patch.object(
+            generate_module, "pad_audio", side_effect=lambda x: x
+        ), mock.patch.object(
+            generate_module, "log_mel_spectrogram", return_value=mx.zeros((128, 8), dtype=mx.float32)
+        ):
+            out = generate_module.generate(
+                model=model,
+                audio_path="dummy.wav",
+                prompt_tokens=[1, 2],
+                n_delay_tokens=0,
+                temperature=0.0,
+                eos_token_id=99,
+                sliding_window=16,
+            )
+
+        self.assertEqual(out, [11, 12, 13])
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/voxmlx/generate.py b/voxmlx/generate.py
index 8b60876..319f5b0 100644
--- a/voxmlx/generate.py
+++ b/voxmlx/generate.py
@@ -73,8 +73,9 @@ def step(token, pos):
 
         y = next_y
 
-    # Check the last token
-    if output_tokens and output_tokens[-1] == eos_token_id:
-        output_tokens = output_tokens[:-1]
+    # Flush last pending token when loop ends without EOS.
+    token_id = y.item()
+    if token_id != eos_token_id:
+        output_tokens.append(token_id)
 
     return output_tokens