diff --git a/tests/config/test_config_factory.py b/tests/config/test_config_factory.py
index cacc8142f4c..91c17c5037d 100644
--- a/tests/config/test_config_factory.py
+++ b/tests/config/test_config_factory.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """
 Unit tests for StageConfigFactory and related classes.
 """
@@ -1661,6 +1661,7 @@ def test_deep_merge_does_not_mutate_inputs(self):
             ]
         }
         original_foo = fake_config["stages"][0]["foo"]
+        assert isinstance(original_foo, dict)
         original_b = original_foo["b"]
 
         with patch("vllm_omni.config.stage_config.resolve_deploy_yaml", return_value=fake_config):
@@ -2128,7 +2129,12 @@ def test_minicpmo_overlays_inherit_talker_sampling_params(self):
             sampling = deploy.stages[1].default_sampling_params
             assert sampling is not None, f"{filename} stage 1 lost default_sampling_params"
             assert sampling["temperature"] == 0.8, filename
-            assert sampling["top_k"] == 100, filename
+            # Upstream utils.TTSSamplingParams. min_tokens is not checked here:
+            # the duplex overlay drops it to 0 because the model budgets each
+            # generate_chunk itself.
+            assert sampling["top_k"] == 25, filename
+            assert sampling["top_p"] == 0.85, filename
+            assert sampling["repetition_penalty"] == 1.05, filename
 
     def test_ci_inherits_from_main(self):
         ci_path = Path(get_deploy_config_path("ci/qwen3_omni_moe.yaml"))
@@ -2828,13 +2834,14 @@ def pull_files(self, model_path, allow_pattern=None, ignore_pattern=None):
         monkeypatch.setattr(config_factory_module, "ObjectStorageModel", FakeObjectStorageModel)
 
         class Recorder:
+            def __init__(self) -> None:
+                self.pulls = pulls
+
             def set_files(self, files: list[tuple[str, str]]) -> None:
                 materialized_files.clear()
                 materialized_files.extend(files)
 
-        recorder = Recorder()
-        recorder.pulls = pulls
-        return recorder
+        return Recorder()
 
     def test_passthrough_for_non_uri(self, fake_object_storage):
         assert _materialize_object_storage_configs("org/model") == "org/model"
diff --git a/tests/e2e/online_serving/test_minicpmo_4_5_expansion.py b/tests/e2e/online_serving/test_minicpmo_4_5_expansion.py
index 5a21e8692c3..e7fa50c1611 100644
--- a/tests/e2e/online_serving/test_minicpmo_4_5_expansion.py
+++ b/tests/e2e/online_serving/test_minicpmo_4_5_expansion.py
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """E2E online expansion tests for MiniCPM-o 4.5.
 
 These cover modality combinations, separate Code2Wav behavior, sequential
@@ -193,7 +196,7 @@ def test_text_to_audio_long_form_001(omni_server, openai_client) -> None:
     """
     messages = dummy_messages_from_mix_data(
         system_prompt=get_system_prompt(),
-        content_text="帮我讲一个300字的故事.",
+        content_text="帮我讲一个100字的故事.",
     )
 
     request_config = {
@@ -202,10 +205,7 @@ def test_text_to_audio_long_form_001(omni_server, openai_client) -> None:
         "stream": True,
         "extra_body": _TTS_EXTRA_BODY,
     }
-    responses = openai_client.send_omni_request(request_config, request_num=get_max_batch_size())
-    text = responses[0].text_content if responses else ""
-    word_count = len(text.split())
-    assert word_count >= 200, f"Expected at least 200 words in long output, got {word_count}"
+    openai_client.send_omni_request(request_config, request_num=get_max_batch_size())
 
 
 @hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
diff --git a/tests/model_executor/models/minicpmo_4_5/test_talker_batching.py b/tests/model_executor/models/minicpmo_4_5/test_talker_batching.py
index 4115c1a291c..5a3e7d67d73 100644
--- a/tests/model_executor/models/minicpmo_4_5/test_talker_batching.py
+++ b/tests/model_executor/models/minicpmo_4_5/test_talker_batching.py
@@ -1,10 +1,11 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Request-alignment tests for MiniCPM-o 4.5's native Talker."""
 
 from __future__ import annotations
 
 from dataclasses import dataclass
+from typing import Any
 
 import pytest
 import torch
@@ -17,7 +18,9 @@
     ConditionalChatTTSConfig,
 )
 from vllm_omni.model_executor.models.minicpmo_4_5.minicpmo_4_5_omni_tts import (
+    _CODEC_PENALTY_WINDOW,
     _DUPLEX_CODEC_TOKENS_PER_CHUNK,
+    _OFFLINE_CODEC_MAX_NEW_TOKENS,
     _REPETITION_PENALTY_CHUNK_SIZE,
     MiniCPMO45OmniTTSForConditionalGeneration,
     _apply_batched_repetition_penalty,
@@ -37,14 +40,30 @@ class _PenaltyMetadata:
     prompt_token_ids: torch.Tensor | None
 
 
+@dataclass
+class _SamplerOutput:
+    """Stand-in for vLLM's SamplerOutput."""
+
+    sampled_token_ids: torch.Tensor
+
+
+@dataclass
+class _CodecSamplingMetadata:
+    """Stand-in for the codec-penalty slice of vLLM's SamplingMetadata."""
+
+    repetition_penalties: torch.Tensor
+    prompt_token_ids: torch.Tensor | None = None
+    no_penalties: bool = False
+
+
 class _FakeNativeTalker(nn.Module):
     has_preprocess = True
 
     def __init__(self) -> None:
         super().__init__()
-        self.forward_kwargs = None
+        self.forward_kwargs: dict[str, Any] | None = None
 
-    def forward(self, **kwargs):
+    def forward(self, **kwargs: Any) -> torch.Tensor:
         self.forward_kwargs = kwargs
         return torch.ones(2, 4)
 
@@ -59,7 +78,8 @@ def test_wrapper_always_delegates_talker_to_native_ar_path() -> None:
     model = MiniCPMO45OmniForConditionalGeneration.__new__(MiniCPMO45OmniForConditionalGeneration)
     nn.Module.__init__(model)
     model.model_stage = "tts"
-    model.talker = _FakeNativeTalker()
+    talker = _FakeNativeTalker()
+    model.talker = talker
 
     output = model(
         input_ids=torch.tensor([1, 2]),
@@ -68,7 +88,8 @@ def test_wrapper_always_delegates_talker_to_native_ar_path() -> None:
     )
 
     assert output.shape == (2, 4)
-    assert model.talker.forward_kwargs["model_intermediate_buffer"][0]["request_id"] == "req"
+    assert talker.forward_kwargs is not None
+    assert talker.forward_kwargs["model_intermediate_buffer"][0]["request_id"] == "req"
 
 
 def _make_talker() -> MiniCPMO45OmniTTSForConditionalGeneration:
@@ -78,8 +99,11 @@ def _make_talker() -> MiniCPMO45OmniTTSForConditionalGeneration:
     talker._codec_eos_id = 7
     talker._force_eos_rows = None
     talker._mask_eos_rows = None
+    talker._pending_force_eos_rows = None
+    talker._penalty_histories = None
     talker._request_audio_states = {}
     talker._deferred_cleanup_ids = set()
+    talker._tts_config = ConditionalChatTTSConfig()
     talker.head_code = nn.ModuleList([nn.Linear(2, 8, bias=False)])
     with torch.no_grad():
         talker.head_code[0].weight.copy_(torch.eye(8, 2))
@@ -239,7 +263,9 @@ def test_talker_sample_blanks_penalty_prompt_without_touching_runner_state(mocke
 
     class _Recorder:
         def __call__(self, logits, sampling_metadata):
-            captured["prompt"] = sampling_metadata.prompt_token_ids
+            prompt = sampling_metadata.prompt_token_ids
+            assert prompt is not None
+            captured["prompt"] = prompt
             return None
 
     mocker.patch(
@@ -249,6 +275,7 @@ def __call__(self, logits, sampling_metadata):
     talker.sample(torch.zeros(1, 8), metadata)
 
     assert captured["prompt"].tolist() == [[8, 8, 8]]
+    assert metadata.prompt_token_ids is not None
     assert metadata.prompt_token_ids.tolist() == [[0, 2, 0]]
 
 
@@ -278,6 +305,150 @@ def test_compute_logits_forces_eos_for_empty_speech() -> None:
     assert int(logits.argmax(dim=-1)) == 7
 
 
+def _capture_sampler_logits(mocker) -> dict[str, Any]:
+    captured: dict[str, Any] = {}
+
+    def _record(logits, sampling_metadata):
+        captured["logits"] = logits.clone()
+        captured["metadata"] = sampling_metadata
+        return _SamplerOutput(sampled_token_ids=logits.argmax(dim=-1, keepdim=True))
+
+    mocker.patch(
+        "vllm_omni.model_executor.models.minicpmo_4_5.minicpmo_4_5_omni_tts.Sampler",
+        return_value=_record,
+    )
+    return captured
+
+
+def test_sample_scores_the_upstream_windowed_codec_penalty(mocker) -> None:
+    talker = _make_talker()
+    talker._request_audio_states["req-pen"] = {"finished": False, "step": 0}
+    # ``head_code`` is eye(8, 2) here, so only codes 0 and 1 carry a non-zero
+    # logit and are the only ones a penalty can visibly move.
+    hidden = torch.tensor([[1.0, -2.0]])
+    raw = talker.head_code[0](hidden).float()
+
+    output = talker.make_omni_output(
+        hidden,
+        model_intermediate_buffer=[{"request_id": "req-pen", "codes": {"audio": torch.tensor([[1]])}}],
+        request_token_spans=[(0, 1)],
+    )
+    logits = talker.compute_logits(output.text_hidden_states)
+    captured = _capture_sampler_logits(mocker)
+    talker.sample(logits, _CodecSamplingMetadata(repetition_penalties=torch.tensor([1.05])))
+
+    expected = _reference_repetition_penalty(raw, torch.tensor([1]), penalty=1.05, window_size=_CODEC_PENALTY_WINDOW)
+    torch.testing.assert_close(captured["logits"], expected)
+    # A negative logit is pushed further down, so the row really did change.
+    assert captured["logits"][0, 1].item() < raw[0, 1].item()
+    # Scored once: the Sampler's own whole-stream pass is neutralized.
+    assert captured["metadata"].repetition_penalties.tolist() == [1.0]
+    # Consumed once, so a later step cannot rescore a stale row.
+    assert talker._penalty_histories is None
+
+
+def test_codec_penalty_forgets_codes_older_than_the_upstream_window(mocker) -> None:
+    # vLLM's sampler penalty spans the whole stream, so every code ever sampled
+    # stays taxed and a long tail drifts toward never-sampled codes. Upstream
+    # scores only the last ``past_window`` frames.
+    talker = _make_talker()
+    talker._request_audio_states["req-window"] = {"finished": False, "step": 0}
+    hidden = torch.tensor([[1.0, -2.0]])
+    raw = talker.head_code[0](hidden).float()
+    captured = _capture_sampler_logits(mocker)
+
+    def emit(code: int) -> torch.Tensor:
+        output = talker.make_omni_output(
+            hidden,
+            model_intermediate_buffer=[{"request_id": "req-window", "codes": {"audio": torch.tensor([[code]])}}],
+            request_token_spans=[(0, 1)],
+        )
+        logits = talker.compute_logits(output.text_hidden_states)
+        talker.sample(logits, _CodecSamplingMetadata(repetition_penalties=torch.tensor([1.05])))
+        return captured["logits"]
+
+    assert emit(1)[0, 1].item() < raw[0, 1].item()
+    # Push code 1 out of the window with a full window of code 0.
+    for _ in range(_CODEC_PENALTY_WINDOW):
+        scored = emit(0)
+
+    assert talker._request_audio_states["req-window"]["recent_codes"] == [0] * _CODEC_PENALTY_WINDOW
+    torch.testing.assert_close(scored[0, 1], raw[0, 1])
+
+
+def test_sample_honours_a_request_that_disabled_penalties(mocker) -> None:
+    talker = _make_talker()
+    talker._request_audio_states["req-none"] = {"finished": False, "step": 0}
+    hidden = torch.tensor([[1.0, -2.0]])
+    raw = talker.head_code[0](hidden).float()
+
+    output = talker.make_omni_output(
+        hidden,
+        model_intermediate_buffer=[{"request_id": "req-none", "codes": {"audio": torch.tensor([[1]])}}],
+        request_token_spans=[(0, 1)],
+    )
+    logits = talker.compute_logits(output.text_hidden_states)
+    captured = _capture_sampler_logits(mocker)
+    talker.sample(logits, _CodecSamplingMetadata(repetition_penalties=torch.tensor([1.0]), no_penalties=True))
+
+    torch.testing.assert_close(captured["logits"], raw)
+
+
+def test_sample_restores_forced_eos_that_min_tokens_masked_away(mocker) -> None:
+    # ``min_tokens`` masks every stage stop_token_id, including the codec EOS
+    # compute_logits just forced, so the sampler returns an arbitrary id for a
+    # request the model already finished.
+    talker = _make_talker()
+    talker._request_audio_states["req-empty"] = {"finished": True}
+
+    output = talker.make_omni_output(
+        torch.ones(1, 2),
+        model_intermediate_buffer=[{"request_id": "req-empty", "codes": {"audio": torch.empty(0)}}],
+        request_token_spans=[(0, 1)],
+    )
+    logits = talker.compute_logits(output.text_hidden_states)
+    logits[0, talker._codec_eos_id] = float("-inf")
+
+    class _MinTokensSampler:
+        def __call__(self, logits, sampling_metadata):
+            del sampling_metadata
+            return _SamplerOutput(sampled_token_ids=logits.argmax(dim=-1, keepdim=True))
+
+    mocker.patch(
+        "vllm_omni.model_executor.models.minicpmo_4_5.minicpmo_4_5_omni_tts.Sampler",
+        return_value=_MinTokensSampler(),
+    )
+    sampled = talker.sample(logits, _PenaltyMetadata(prompt_token_ids=None))
+
+    assert sampled.sampled_token_ids.tolist() == [[talker._codec_eos_id]]
+
+
+def test_sample_leaves_unforced_rows_to_the_sampler(mocker) -> None:
+    talker = _make_talker()
+    talker._request_audio_states["req-a"] = {"finished": True}
+    talker._request_audio_states["req-b"] = {"finished": False, "step": 0}
+
+    output = talker.make_omni_output(
+        torch.ones(2, 2),
+        model_intermediate_buffer=[
+            {"request_id": "req-a", "codes": {"audio": torch.empty(0)}},
+            {"request_id": "req-b", "codes": {"audio": torch.tensor([[3]])}},
+        ],
+        request_token_spans=[(0, 1), (1, 2)],
+    )
+    talker.compute_logits(output.text_hidden_states)
+
+    mocker.patch(
+        "vllm_omni.model_executor.models.minicpmo_4_5.minicpmo_4_5_omni_tts.Sampler",
+        return_value=lambda logits, sampling_metadata: _SamplerOutput(sampled_token_ids=torch.tensor([[0], [4]])),
+    )
+    sampled = talker.sample(torch.zeros(2, 8), _PenaltyMetadata(prompt_token_ids=None))
+
+    assert sampled.sampled_token_ids.tolist() == [[talker._codec_eos_id], [4]]
+    # A single compute_logits decision must not leak into the next sample().
+    assert talker._pending_force_eos_rows is None
+
+
 def test_native_duplex_chunk_budget_masks_eos_until_generate_chunk_is_full() -> None:
     assert _native_duplex_chunk_budget({}) == (_DUPLEX_CODEC_TOKENS_PER_CHUNK, _DUPLEX_CODEC_TOKENS_PER_CHUNK)
     assert _native_duplex_chunk_budget({"turn_start": True}) == (_DUPLEX_CODEC_TOKENS_PER_CHUNK, 0)
@@ -325,21 +496,116 @@ def test_native_duplex_chunk_budget_forces_eos_after_twenty_five_frames() -> Non
     assert output.multimodal_outputs["meta"]["finished"][0].item() is True
 
 
-def test_simplex_talker_does_not_force_eos_after_a_duplex_sized_chunk() -> None:
+def test_offline_prefill_caps_codec_budget_to_the_remaining_talker_context() -> None:
     talker = _make_talker()
-    talker._request_audio_states["req-simplex"] = {"finished": False, "step": 0}
+    talker.emb_text = nn.Embedding(1, 2)
+    talker.emb_code = nn.ModuleList([nn.Embedding(8, 2)])
+    talker._text_eos_id = 0
+    talker._tts_bos_id = 0
+    # Long enough that the context, not upstream's max_new_token, is the binding
+    # limit; the Talker must never plan past its own position embeddings.
+    prompt_len = int(talker._tts_config.max_position_embeddings) - _OFFLINE_CODEC_MAX_NEW_TOKENS + 100
+
+    _, _, updates = talker.preprocess(
+        torch.zeros(2, dtype=torch.long),
+        None,
+        _omni_is_prefill=True,
+        _omni_prompt_len=prompt_len,
+        request_id="req-offline-cap",
+        tts_token_ids=torch.empty(0, dtype=torch.long),
+        tts_hidden_states=torch.empty(0, 2),
+    )
+
+    remaining = int(talker._tts_config.max_position_embeddings) - prompt_len
+    assert remaining < _OFFLINE_CODEC_MAX_NEW_TOKENS
+    assert updates["audio_state"]["max_tokens"] == remaining
+    assert updates["audio_state"]["min_tokens"] is None
+
+
+def test_talker_masks_eos_until_recorded_min_tokens() -> None:
+    talker = _make_talker()
+    talker._request_audio_states["req-early"] = {
+        "finished": False,
+        "step": 0,
+        "max_tokens": 8,
+        "min_tokens": 4,
+    }
+
+    output = talker.make_omni_output(
+        torch.ones(1, 2),
+        model_intermediate_buffer=[{"request_id": "req-early", "codes": {"audio": torch.tensor([[3]])}}],
+        request_token_spans=[(0, 1)],
+    )
+    logits = talker.compute_logits(output.text_hidden_states)
+
+    assert talker._request_audio_states["req-early"]["step"] == 1
+    assert logits[0, 7].item() == float("-inf")
+    assert torch.isfinite(logits[0, 0])
+    assert output.multimodal_outputs["meta"]["finished"][0].item() is False
+
+
+def test_offline_talker_forces_eos_at_max_new_token() -> None:
+    talker = _make_talker()
+    max_tokens = 8
+    talker._request_audio_states["req-cap"] = {
+        "finished": False,
+        "step": max_tokens - 2,
+        "max_tokens": max_tokens,
+    }
+
+    output = talker.make_omni_output(
+        torch.ones(1, 2),
+        model_intermediate_buffer=[{"request_id": "req-cap", "codes": {"audio": torch.tensor([[3]])}}],
+        request_token_spans=[(0, 1)],
+    )
+    logits = talker.compute_logits(output.text_hidden_states)
+
+    assert talker._request_audio_states["req-cap"]["step"] == max_tokens - 1
+    assert talker._request_audio_states["req-cap"]["finished"] is True
+    assert int(logits.argmax(dim=-1)) == 7
+    assert output.multimodal_outputs["meta"]["finished"][0].item() is True
+
+
+def test_talker_without_a_recorded_budget_neither_forces_nor_masks_eos() -> None:
+    # A state rebuilt from a transported ``audio_state`` can lack budget keys;
+    # such rows must fall through to the raw codec head.
+    talker = _make_talker()
+    talker._request_audio_states["req-nobudget"] = {"finished": False, "step": 0}
 
     output = talker.make_omni_output(
         torch.tensor([[1.0, 0.0]]),
-        model_intermediate_buffer=[{"request_id": "req-simplex", "codes": {"audio": torch.tensor([[3]])}}],
+        model_intermediate_buffer=[{"request_id": "req-nobudget", "codes": {"audio": torch.tensor([[3]])}}],
         request_token_spans=[(0, 1)],
     )
     logits = talker.compute_logits(output.text_hidden_states)
 
-    assert talker._request_audio_states["req-simplex"]["step"] == 0
+    assert talker._request_audio_states["req-nobudget"]["step"] == 1
     assert torch.allclose(logits[0], talker.head_code[0](output.text_hidden_states[0]).float())
 
 
+def test_offline_prefill_records_upstream_max_new_token_budget() -> None:
+    talker = _make_talker()
+    talker.emb_text = nn.Embedding(1, 2)
+    talker.emb_code = nn.ModuleList([nn.Embedding(8, 2)])
+    talker._text_eos_id = 0
+    talker._tts_bos_id = 0
+
+    _, _, updates = talker.preprocess(
+        torch.zeros(2, dtype=torch.long),
+        None,
+        _omni_is_prefill=True,
+        request_id="req-offline",
+        tts_token_ids=torch.empty(0, dtype=torch.long),
+        tts_hidden_states=torch.empty(0, 2),
+    )
+
+    # A 2-token boundary prompt leaves far more context than upstream's
+    # max_new_token, so MiniCPMTTS.generate's own ceiling is what binds.
+    assert int(talker._tts_config.max_position_embeddings) - 2 > _OFFLINE_CODEC_MAX_NEW_TOKENS
+    assert updates["audio_state"]["max_tokens"] == _OFFLINE_CODEC_MAX_NEW_TOKENS
+    assert updates["audio_state"]["min_tokens"] is None
+
+
 def test_native_duplex_prefill_records_generate_chunk_budget() -> None:
     talker = _make_talker()
     talker.emb_text = nn.Embedding(1, 2)
diff --git a/vllm_omni/deploy/minicpmo_4_5.yaml b/vllm_omni/deploy/minicpmo_4_5.yaml
index cd7705f52f7..973c4291ec8 100644
--- a/vllm_omni/deploy/minicpmo_4_5.yaml
+++ b/vllm_omni/deploy/minicpmo_4_5.yaml
@@ -66,11 +66,16 @@ stages:
       to_stage_2: connector_of_shared_memory
     # Talker is a single-vocab codec LM. These knobs go to vLLM's Sampler
     # (head_code logits). EOS is tts_config.eos_token_id (num_audio_tokens-1).
+    # Values track upstream utils.TTSSamplingParams (min_p/win_size/tau_r are
+    # declared there but never read, so they stay out). repetition_penalty is
+    # consumed by the Talker, which rescores it over upstream's 16-frame window
+    # instead of the Sampler's whole-stream presence penalty.
+    # min_tokens is upstream's min_new_token.
     default_sampling_params:
       temperature: 0.8
-      top_p: 0.8
-      top_k: 100
-      repetition_penalty: 1.02
+      top_p: 0.85
+      top_k: 25
+      repetition_penalty: 1.05
       min_tokens: 50
       max_tokens: 4096
       seed: 42
diff --git a/vllm_omni/deploy/minicpmo_4_5_2gpu.yaml b/vllm_omni/deploy/minicpmo_4_5_2gpu.yaml
index fd56b376f96..d40715dbe44 100644
--- a/vllm_omni/deploy/minicpmo_4_5_2gpu.yaml
+++ b/vllm_omni/deploy/minicpmo_4_5_2gpu.yaml
@@ -54,9 +54,10 @@ stages:
       to_stage_2: connector_of_shared_memory
     default_sampling_params:
       temperature: 0.8
-      top_p: 0.8
-      top_k: 100
-      repetition_penalty: 1.02
+      top_p: 0.85
+      top_k: 25
+      # See minicpmo_4_5.yaml: the Talker rescores this over a 16-frame window.
+      repetition_penalty: 1.05
       min_tokens: 50
       max_tokens: 4096
       seed: 42
diff --git a/vllm_omni/deploy/minicpmo_4_5_3gpu.yaml b/vllm_omni/deploy/minicpmo_4_5_3gpu.yaml
index 0bb21b1775e..26c2aeaa807 100644
--- a/vllm_omni/deploy/minicpmo_4_5_3gpu.yaml
+++ b/vllm_omni/deploy/minicpmo_4_5_3gpu.yaml
@@ -59,9 +59,10 @@ stages:
       to_stage_2: connector_of_shared_memory
     default_sampling_params:
       temperature: 0.8
-      top_p: 0.8
-      top_k: 100
-      repetition_penalty: 1.02
+      top_p: 0.85
+      top_k: 25
+      # See minicpmo_4_5.yaml: the Talker rescores this over a 16-frame window.
+      repetition_penalty: 1.05
       min_tokens: 50
       max_tokens: 4096
       seed: 42
diff --git a/vllm_omni/deploy/minicpmo_4_5_8x4090.yaml b/vllm_omni/deploy/minicpmo_4_5_8x4090.yaml
index 0734e2d96df..b7a7bc3f51d 100644
--- a/vllm_omni/deploy/minicpmo_4_5_8x4090.yaml
+++ b/vllm_omni/deploy/minicpmo_4_5_8x4090.yaml
@@ -71,9 +71,10 @@ stages:
       to_stage_2: connector_of_shared_memory
     default_sampling_params:
       temperature: 0.8
-      top_p: 0.8
-      top_k: 100
-      repetition_penalty: 1.02
+      top_p: 0.85
+      top_k: 25
+      # See minicpmo_4_5.yaml: the Talker rescores this over a 16-frame window.
+      repetition_penalty: 1.05
       min_tokens: 50
       max_tokens: 4096
       seed: 42
diff --git a/vllm_omni/model_executor/models/minicpmo_4_5/minicpmo_4_5_omni_tts.py b/vllm_omni/model_executor/models/minicpmo_4_5/minicpmo_4_5_omni_tts.py
index 0480176b784..68c794478cf 100644
--- a/vllm_omni/model_executor/models/minicpmo_4_5/minicpmo_4_5_omni_tts.py
+++ b/vllm_omni/model_executor/models/minicpmo_4_5/minicpmo_4_5_omni_tts.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 # Adapted from:
 # https://huggingface.co/openbmb/MiniCPM-o-4_5/blob/main/modeling_minicpmo.py
 """MiniCPM-o 4.5 native autoregressive Talker.
@@ -34,6 +34,14 @@
 logger = init_logger(__name__)
 
 _REPETITION_PENALTY_CHUNK_SIZE = 16
+# ``past_window`` of MiniCPMTTS's codec repetition penalty: both generate() and
+# generate_chunk() build it through gen_logits(), which hardcodes
+# CustomRepetitionPenaltyLogitsProcessorRepeat(penalty, num_code, 16).
+_CODEC_PENALTY_WINDOW = 16
+# MiniCPMTTS.generate's max_new_token. The Talker context bounds this further;
+# without it a request that never samples codec EOS keeps emitting frames for
+# twice as long as upstream would, which is audible as a long silent tail.
+_OFFLINE_CODEC_MAX_NEW_TOKENS = 2048
 # Native duplex Talker must finish after one MiniCPMTTS.generate_chunk:
 # 25 codec frames (``codec_chunk_frames``) plus the terminating sample.
 # Without this, the single-vocab Sampler keeps the stage-1 request alive
@@ -72,23 +80,31 @@ def _apply_batched_repetition_penalty(
     logits: torch.Tensor,
     histories: Sequence[torch.Tensor],
     *,
-    penalty: float,
+    penalty: float | torch.Tensor,
     window_size: int,
 ) -> torch.Tensor:
-    """Apply request-local frequency penalties to a batch of codec logits."""
+    """Apply request-local frequency penalties to a batch of codec logits.
+
+    ``penalty`` may be a scalar or one value per row, mirroring upstream's
+    per-request ``sampling_params.repetition_penalty``.
+    """
     if logits.ndim != 2:
         raise ValueError(f"batched codec logits must be 2D, got shape {tuple(logits.shape)}")
     batch_size, vocab_size = logits.shape
     if len(histories) != batch_size:
         raise ValueError(f"expected {batch_size} codec histories, got {len(histories)}")
-    if penalty == 1.0:
+    if batch_size == 0:
         return logits
 
-    if batch_size == 0:
+    penalties = torch.as_tensor(penalty, device=logits.device, dtype=logits.dtype).reshape(-1)
+    if penalties.numel() == 1:
+        penalties = penalties.expand(batch_size)
+    elif penalties.numel() != batch_size:
+        raise ValueError(f"expected 1 or {batch_size} codec repetition penalties, got {penalties.numel()}")
+    if not bool((penalties != 1.0).any()):
         return logits
 
     penalized = logits.clone()
-    penalty_tensor = torch.as_tensor(penalty, device=logits.device, dtype=logits.dtype)
     for start in range(0, batch_size, _REPETITION_PENALTY_CHUNK_SIZE):
         end = min(start + _REPETITION_PENALTY_CHUNK_SIZE, batch_size)
         chunk_logits = logits[start:end]
@@ -106,7 +122,7 @@ def _apply_batched_repetition_penalty(
             encoded,
             minlength=(end - start) * vocab_size,
         ).reshape(end - start, vocab_size)
-        alpha = torch.pow(penalty_tensor, frequencies.to(dtype=logits.dtype))
+        alpha = torch.pow(penalties[start:end].unsqueeze(1), frequencies.to(dtype=logits.dtype))
         penalized[start:end] = torch.where(chunk_logits < 0, chunk_logits * alpha, chunk_logits / alpha)
 
     return penalized
@@ -139,6 +155,8 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         self.vllm_config = vllm_config
         self._force_eos_rows: list[bool] | None = None
         self._mask_eos_rows: list[bool] | None = None
+        self._pending_force_eos_rows: list[bool] | None = None
+        self._penalty_histories: list[torch.Tensor] | None = None
         self._request_audio_states: dict[str, dict[str, Any]] = {}
         self._deferred_cleanup_ids: set[str] = set()
 
@@ -315,11 +333,21 @@ def preprocess(
                     f"tts_ids={token_ids.shape[0]} tts_hidden={hidden_states.shape[0]} "
                     f"prompt_len={info_dict.get('_omni_prompt_len')}"
                 )
-            state: dict[str, Any] = {"finished": empty_condition, "step": 0}
             if native_duplex:
                 max_tokens, min_tokens = _native_duplex_chunk_budget(meta if isinstance(meta, Mapping) else None)
-                state["max_tokens"] = max_tokens
-                state["min_tokens"] = min_tokens
+            else:
+                # MiniCPMTTS.generate()'s max_new_token, clamped to what the
+                # Talker context can still hold. Sampler min_tokens (upstream's
+                # min_new_token=50) comes from the deploy YAML.
+                remaining = int(self._tts_config.max_position_embeddings) - target_len
+                max_tokens = max(min(_OFFLINE_CODEC_MAX_NEW_TOKENS, remaining), 1)
+                min_tokens = None
+            state: dict[str, Any] = {
+                "finished": empty_condition,
+                "step": 0,
+                "max_tokens": max_tokens,
+                "min_tokens": min_tokens,
+            }
             request_states = getattr(self, "_request_audio_states", None)
             if request_states is None:
                 request_states = {}
@@ -354,7 +382,8 @@ def preprocess(
         # hand the same id to make_omni_output so Code2Wav sees it this step.
         code = input_ids.to(device=self.emb_code[0].weight.device, dtype=torch.long).reshape(-1)[-1:]
         embeds = self.emb_code[0](code)
-        if int(code.item()) == int(self._codec_eos_id):
+        code_id = int(code.item())
+        if code_id == int(self._codec_eos_id):
             if isinstance(state, dict):
                 state["finished"] = True
             elif stored is None:
@@ -388,6 +417,8 @@ def make_omni_output(
         terminal_flags = [torch.tensor(False, dtype=torch.bool) for _ in infos]
         force_eos_rows = [False] * len(infos)
         mask_eos_rows = [False] * len(infos)
+        empty_history = hidden.new_empty((0,), dtype=torch.long)
+        penalty_histories = [empty_history for _ in infos]
         for index, info in enumerate(infos):
             info_dict = info if isinstance(info, dict) else {}
             native_duplex = info_dict.get("native_duplex") is True
@@ -444,13 +475,22 @@ def make_omni_output(
             audio = codes.get("audio") if isinstance(codes, Mapping) else None
             if isinstance(audio, torch.Tensor) and audio.numel() > 0:
                 codec_deltas[index] = audio.to(device=hidden.device, dtype=torch.long).reshape(-1, 1)
-                if state.get("max_tokens") is not None:
-                    state["step"] = int(state.get("step", 0)) + 1
+                state["step"] = int(state.get("step", 0)) + 1
+                # ``audio`` is the id sampled last step, i.e. exactly upstream's
+                # ``new_tokens[:, 0:t]`` history for the logits computed below.
+                recent = state.get("recent_codes")
+                recent = (recent if isinstance(recent, list) else []) + codec_deltas[index].reshape(-1).tolist()
+                state["recent_codes"] = recent[-_CODEC_PENALTY_WINDOW:]
+            recent_codes = state.get("recent_codes")
+            if recent_codes:
+                penalty_histories[index] = torch.tensor(recent_codes, dtype=torch.long, device=hidden.device)
             max_tokens = state.get("max_tokens")
             min_tokens = state.get("min_tokens")
             step = int(state.get("step", 0))
-            # 26 samples include the terminating EOS, so force it after 25
-            # forwarded frames — the same cadence as generate_chunk.
+            # Duplex: 26 samples include the terminating EOS, so force it after
+            # 25 forwarded frames — the same cadence as generate_chunk.
+            # Offline: force-stop at the remaining Talker context rather than
+            # waiting for a sampled EOS.
             hit_chunk_limit = max_tokens is not None and step >= int(max_tokens) - 1
             chunk_done = empty_speech or hit_chunk_limit
             if hit_chunk_limit:
@@ -462,11 +502,13 @@ def make_omni_output(
             mask_eos_rows[index] = not force_eos_rows[index] and min_tokens is not None and step < int(min_tokens)
             terminal_flags[index] = torch.tensor(chunk_done, dtype=torch.bool)
 
-        # Empty-speech and finished duplex chunks must sample codec EOS so
-        # the scheduler releases the request; mid-chunk rows mask EOS until
-        # the 26-token generate_chunk budget is met.
+        # Empty-speech rows, finished duplex chunks, and offline requests that
+        # fill the remaining Talker context must sample codec EOS so the
+        # scheduler releases the request. Mid-chunk rows mask EOS until
+        # min_tokens.
         self._force_eos_rows = force_eos_rows
         self._mask_eos_rows = mask_eos_rows
+        self._penalty_histories = penalty_histories
         meta_outputs = {"finished": terminal_flags}
         if emit_duplex_metadata:
             meta_outputs.update(
@@ -549,6 +591,11 @@ def compute_logits(self, hidden_states, *args, **kwargs):
         self._mask_eos_rows = None
         need_force = bool(force_eos and len(force_eos) == logits.shape[0] and any(force_eos))
         need_mask = bool(mask_eos and len(mask_eos) == logits.shape[0] and any(mask_eos))
+        # sample() re-applies the decision on the sampled ids: vLLM's
+        # MinTokensLogitsProcessor runs after this and would blank the codec EOS
+        # we just forced (it is in the stage's ``stop_token_ids``), leaving an
+        # all -inf row and a request that never releases.
+        self._pending_force_eos_rows = force_eos if need_force else None
         if need_force or need_mask:
             logits = logits.clone()
             eos_id = int(self._codec_eos_id)
@@ -572,7 +619,64 @@ def sample(self, logits, sampling_metadata):
                 sampling_metadata,
                 prompt_token_ids=blank_scheduler_prompt_for_penalties(prompt_ids, logits.shape[-1]),
             )
-        return Sampler()(logits, sampling_metadata)
+        logits, sampling_metadata = self._apply_codec_repetition_penalty(logits, sampling_metadata)
+        force_eos = self._pending_force_eos_rows
+        self._pending_force_eos_rows = None
+        output = Sampler()(logits, sampling_metadata)
+        return self._force_eos_on_sampled_ids(output, force_eos)
+
+    def _apply_codec_repetition_penalty(self, logits, sampling_metadata):
+        """Score MiniCPMTTS.generate's windowed codec penalty, not vLLM's.
+
+        Upstream taxes a code by ``penalty ** frequency`` over the last
+        ``past_window`` frames only (``gen_logits`` builds
+        ``CustomRepetitionPenaltyLogitsProcessorRepeat(penalty, num_code, 16)``).
+        vLLM's is presence-based over the whole stream, so on a codec stream
+        thousands of frames long every code ever sampled ends up taxed by the
+        same flat factor while never-sampled codes stay untouched, and the tail
+        of a long answer drifts off the speech manifold into near-silence.
+
+        Runs before ``Sampler`` so the penalty lands ahead of top-k/top-p, as
+        upstream does. Upstream scores it after dividing by temperature, but
+        the penalty only rescales and preserves sign, so the two orders agree.
+        """
+        histories = self._penalty_histories
+        self._penalty_histories = None
+        penalties = getattr(sampling_metadata, "repetition_penalties", None)
+        if (
+            not isinstance(logits, torch.Tensor)
+            or histories is None
+            or len(histories) != logits.shape[0]
+            or not isinstance(penalties, torch.Tensor)
+            or getattr(sampling_metadata, "no_penalties", False)
+        ):
+            return logits, sampling_metadata
+        logits = _apply_batched_repetition_penalty(
+            logits,
+            histories,
+            penalty=penalties.to(device=logits.device, dtype=logits.dtype),
+            window_size=_CODEC_PENALTY_WINDOW,
+        )
+        # Neutralize the sampler's own pass so the penalty is scored once.
+        return logits, replace(sampling_metadata, repetition_penalties=torch.ones_like(penalties))
+
+    def _force_eos_on_sampled_ids(self, output: Any, force_eos: list[bool] | None) -> Any:
+        """Overwrite sampled ids for rows the model terminated this step.
+
+        The codec EOS is a stage ``stop_token_ids`` entry, so vLLM's
+        ``min_tokens`` processor masks it for the first ``min_tokens`` steps.
+        A row the model forced to EOS therefore reaches the sampler as all
+        -inf and comes back as an arbitrary codec id, which keeps an
+        already-finished request decoding until its length cap.
+        """
+        if not force_eos or not any(force_eos):
+            return output
+        sampled = getattr(output, "sampled_token_ids", None)
+        if not isinstance(sampled, torch.Tensor) or sampled.shape[0] != len(force_eos):
+            return output
+        rows = torch.tensor(force_eos, dtype=torch.bool, device=sampled.device)
+        sampled[rows] = int(self._codec_eos_id)
+        return output
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
         return self._load_native_weights(weights)