docs: clarify model checkpoint and serving choices

docs/README.md

@@ -33,7 +33,8 @@
 | Distinguish speaker vectors, anonymous labels and emotion tags | [Speakers and emotion tags](speaker_emotion.md) / [中文](speaker_emotion_zh.md) |
 | Transcribe and diarize with third-party MOSS | [MOSS](moss_transcribe_diarize.md) / [中文](moss_transcribe_diarize_zh.md) |
 | Accelerate with the FunASR vLLM split engine | [vLLM](vllm_guide.md) / [中文](vllm_guide_zh.md) |
-| Evaluate native vLLM serving | [Validation record](vllm_native_funasr_validation.md) |
+| Evaluate official native vLLM serving | [Official checkpoint validation](vllm_official_native_validation.md) / [中文](vllm_official_native_validation_zh.md) |
+| Inspect the historical community native vLLM run (2026-08-13) | [Historical community record](vllm_native_funasr_validation.md) |
 | Deploy llama.cpp, TensorRT, Docker or Kubernetes | [Deployment manuals](https://www.funasr.com/en/deploy/) |
 | Choose a service entry point and runtime SDK | [Runtime guide](../runtime/readme.md) / [中文](../runtime/readme_cn.md) |
 | Select a development or serving container | [Docker](installation/docker.md) / [中文](installation/docker_zh.md) |

docs/model_selection.md

@@ -42,7 +42,7 @@ Switch to Paraformer when your workload is Mandarin-only and you want character-
 | Fast multilingual private transcription | SenseVoice-Small | Strong default with ASR, emotion tags, audio event tags, and CPU viability. | [README quick start](../README.md#quick-start) |
 | Mandarin production ASR | Paraformer-Large | Mature Chinese ASR path with VAD and punctuation. | [Tutorial](./tutorial/README.md) |
 | English-only route in the OpenAI API example | `paraformer-en` alias | Smaller English route for API compatibility checks. | [OpenAI API example](../examples/openai_api/) |
-| LLM-based ASR or Chinese/English/Japanese + dialect experiments | Fun-ASR-Nano | LLM-based model path; use vLLM when decoder throughput matters. | [vLLM guide](./vllm_guide.md) |
+| LLM-based ASR or Chinese/English/Japanese + dialect experiments | Fun-ASR-Nano | Evaluate the Python path first; split-engine and native vLLM use different loading contracts. | [Choose a vLLM path](#vllm-checkpoint-paths) |
 | Offline long-form ASR with anonymous diarization | MOSS-Transcribe-Diarize | One offline request returns transcription, timestamps, and per-recording anonymous speaker labels; it does not identify known people and needs no external VAD or speaker model. | [MOSS deployment guide](./moss_transcribe_diarize.md) |
 | Live captions or call-center streams | Runtime WebSocket service | Designed for long-lived streaming sessions and partial results. | [Runtime service docs](../runtime/readme.md) |
 | Batch archive processing | SenseVoice-Small or Paraformer-Large | Stable offline transcription path; caller owns manifests, retries, and logs. | [Batch ASR example](../examples/batch_asr_improved.py) |
@@ -56,6 +56,13 @@ The `examples/openai_api` server exposes short aliases so application teams do n
 - **`paraformer`** uses `paraformer-zh` with VAD and punctuation for a Mandarin-oriented route.
 - **`paraformer-en`** uses `paraformer-en` with VAD for English transcription in OpenAI-style clients.
 - **`fun-asr-nano`** uses `FunAudioLLM/Fun-ASR-Nano-2512` for evaluating Chinese, English, Japanese, and Chinese dialect/accent coverage. Select a compatible runtime when evaluating vLLM acceleration.
+- **`moss-transcribe-diarize`** uses the third-party `OpenMOSS-Team/MOSS-Transcribe-Diarize` model for offline transcription and anonymous per-recording speaker labels. Prepare its separate dependencies and reviewed remote code using the [MOSS guide](./moss_transcribe_diarize.md); request `verbose_json` for structured segments. It does not require an external VAD/speaker model and does not identify known people.
+
+These aliases describe [the example server](../examples/openai_api/server.py),
+which loads `AutoModel`. They do not configure native vLLM or automatically select
+`AutoModelVLLM`. The packaged `funasr-server` has a separate loader and backend
+selection; do not copy an alias or a performance result between services without
+checking the corresponding [HTTP guide](../examples/openai_api/README.md).
 
 The example HTTP service cleans both top-level `text` and segment `text` in
 `verbose_json`; that format does not restore emotion/event tags. If you need
@@ -82,10 +89,27 @@ For SDK, JavaScript, workflow, Postman, OpenAPI, Docker, and Kubernetes paths, s
 | Repeatable local container demo | Docker Compose API | CPU-first smoke test; adapt the image before using CUDA. |
 | Internal cluster service | Kubernetes API template | Private `ClusterIP`, persistent model cache, `/health` probes, and port-forward smoke test. |
 | Live audio | Runtime WebSocket service | Validate chunk size, VAD, endpointing, reconnects, and client backpressure with real audio. |
-| LLM-based ASR throughput | vLLM path for Fun-ASR-Nano | vLLM accelerates autoregressive decoding; it does not apply to non-autoregressive Paraformer. |
+| LLM-based ASR throughput | Choose split-engine or native vLLM below | Match the checkpoint, loading API and tested environment; this is not a Paraformer backend. |
 
 See the [deployment matrix](./deployment_matrix.md) when you are choosing between these paths.
 
+<a id="vllm-checkpoint-paths"></a>
+
+## Choose the vLLM Checkpoint and Interface
+
+| Path | Checkpoint and interface | Read next |
+| --- | --- | --- |
+| FunASR split-engine | Base `FunAudioLLM/Fun-ASR-Nano-2512` assets through `AutoModelVLLM`; FunASR handles the audio side and vLLM the decoder. | [Split-engine preparation and limits](./vllm_guide.md) |
+| Official native vLLM | Converted `FunAudioLLM/Fun-ASR-Nano-2512-vllm` snapshot through vLLM's native model implementation and `/v1/audio/transcriptions`. Not an `AutoModelVLLM` load. | [Official functional validation](./vllm_official_native_validation.md) |
+| Historical community native vLLM | Community `allendou/Fun-ASR-Nano-2512-vllm`, tested on 2026-08-13. Its timings belong to that checkpoint and environment. | [Historical community record](./vllm_native_funasr_validation.md) |
+
+The official record pins a model revision and an existing environment; it is not
+a clean-install recipe, a sustained-load benchmark, or proof of `/v1/realtime`
+streaming. Do not reuse the historical community timings for the official model.
+For MOSS, follow its own deployment guide: the Nano checkpoints and validation
+above do not establish MOSS runtime compatibility. Choose model, checkpoint,
+interface and environment together before testing your own workload.
+
 ## Benchmark before committing
 
 Do not choose a model from a single clean demo file. Use a small representative set first:

docs/model_selection_ja.md

@@ -29,7 +29,7 @@ result = model.generate(input="meeting.wav")
 | 高速な多言語プライベート文字起こし | SenseVoice-Small | ASR、感情タグ、音声イベントタグ、CPU/GPU の扱いやすさがそろった標準ルート。 | [README quick start](../README_ja.md#クイックスタート) |
 | 中国語中心の本番 ASR | Paraformer-Large | VAD と句読点復元を組み合わせた成熟した中国語 ASR ルート。 | [Tutorial](./tutorial/README.md) |
 | OpenAI API 例の英語ルート | `paraformer-en` alias | OpenAI-style client で互換性を確認しやすい軽量な英語ルート。 | [OpenAI API example](../examples/openai_api/README_ja.md) |
-| LLM-based ASR や中英日 + 中国語方言・地域アクセントの評価 | Fun-ASR-Nano | LLM-based モデル。decoder throughput が重要なら vLLM を使います。 | [vLLM guide](./vllm_guide.md) |
+| LLM-based ASR や中英日 + 中国語方言・地域アクセントの評価 | Fun-ASR-Nano | Python で評価してから、checkpoint と interface に合わせて vLLM の経路を選びます。 | [vLLM の経路](#vllm-checkpoint-paths) |
 | オフライン長時間 ASR と匿名話者ラベル | MOSS-Transcribe-Diarize | 1 回のオフライン request で文字起こし、timestamps、録音内の匿名話者ラベルを返します。既知人物の識別ではなく、外部 VAD / speaker model も不要です。 | [MOSS deployment guide](./moss_transcribe_diarize.md) |
 | ライブ字幕やコールセンターストリーム | Runtime WebSocket service | 長時間接続、部分結果、エンドポイント検出に向いたランタイム。 | [Runtime service docs](../runtime/readme.md) |
 | Whisper / cloud ASR からの移行 | SenseVoice-Small で baseline を作り、必要に応じて比較 | まず強い標準ルートで評価してから、用途別に詰めるのが安全です。 | [Migration guide](./migration_from_whisper.md) |
@@ -42,6 +42,12 @@ result = model.generate(input="meeting.wav")
 - **`paraformer`**: `paraformer-zh` に VAD と句読点復元を組み合わせた中国語向けの経路です。
 - **`paraformer-en`**: `paraformer-en` と VAD を使う、OpenAI-style client 向けの英語文字起こしです。
 - **`fun-asr-nano`**: `FunAudioLLM/Fun-ASR-Nano-2512` による中英日・中国語方言/地域アクセントの評価経路です。vLLM acceleration を試す場合は互換性のある runtime を選んでください。
+- **`moss-transcribe-diarize`**: 第三者の `OpenMOSS-Team/MOSS-Transcribe-Diarize` によるオフライン文字起こしと録音内の匿名話者ラベルです。[MOSS guide（英語）](./moss_transcribe_diarize.md)で専用の依存関係と remote code を確認し、構造化 segment には `verbose_json` を指定します。外部 VAD / speaker model は不要で、既知人物の識別ではありません。
+
+ここでの alias は `AutoModel` を読み込む[サンプル server](../examples/openai_api/server.py)のものです。
+native vLLM や `AutoModelVLLM` を自動選択する設定ではありません。
+パッケージの `funasr-server` は別の loader / backend 選択を持つため、
+サービス間で alias や性能結果をそのまま流用しないでください。
 
 この HTTP サンプルはトップレベルの `text` と `verbose_json` の各 segment の
 `text` を整形するため、形式を変えても感情/イベントタグは復元されません。
@@ -57,6 +63,21 @@ python examples/openai_api/smoke_test.py --base-url http://localhost:8000 --mode
 
 SDK、JavaScript、workflow、Postman、OpenAPI、Docker、Kubernetes は [OpenAI API example](../examples/openai_api/README_ja.md) から始めてください。
 
+<a id="vllm-checkpoint-paths"></a>
+
+## vLLM の checkpoint と interface
+
+| 経路 | checkpoint と interface | 次の資料 |
+| --- | --- | --- |
+| FunASR split-engine | 基本の `FunAudioLLM/Fun-ASR-Nano-2512` を `AutoModelVLLM` で読み込み、音声側は FunASR、decoder は vLLM が処理します。 | [Split-engine（英語）](./vllm_guide.md) |
+| 公式 native vLLM | 変換済み `FunAudioLLM/Fun-ASR-Nano-2512-vllm` を vLLM の native 実装で読み込み、`/v1/audio/transcriptions` を利用します。`AutoModelVLLM` ではありません。 | [公式機能検証（英語）](./vllm_official_native_validation.md) |
+| 過去の community native vLLM | `allendou/Fun-ASR-Nano-2512-vllm`、2026-08-13 の検証です。時間計測は当時の checkpoint と環境に限定されます。 | [過去の community 記録](./vllm_native_funasr_validation.md) |
+
+公式記録は固定 revision と既存環境での機能検証であり、新規インストール手順、
+持続負荷の benchmark、`/v1/realtime` の streaming 検証ではありません。
+過去の community 計測を公式モデルに付け替えないでください。
+MOSS は専用ガイドに従ってください。Nano の checkpoint と検証は MOSS の互換性を証明しません。
+
 ## ベンチマークしてから決める
 
 きれいな demo 音声 1 つだけでモデルを決めないでください。まず小さな代表セットで確認します。

docs/model_selection_ko.md

@@ -29,7 +29,7 @@ result = model.generate(input="meeting.wav")
 | 빠른 다국어 프라이빗 전사 | SenseVoice-Small | ASR, 감정 태그, 음성 이벤트 태그, CPU/GPU 사용성이 균형 잡힌 기본 경로입니다. | [README quick start](../README_ko.md#빠른-시작) |
 | 중국어 중심 프로덕션 ASR | Paraformer-Large | VAD와 문장부호 복원을 함께 쓰는 성숙한 중국어 ASR 경로입니다. | [Tutorial](./tutorial/README.md) |
 | OpenAI API 예제의 영어 경로 | `paraformer-en` alias | OpenAI-style client에서 호환성을 확인하기 쉬운 가벼운 영어 경로입니다. | [OpenAI API example](../examples/openai_api/README_ko.md) |
-| LLM-based ASR 또는 중영일 + 중국어 방언/지역 억양 평가 | Fun-ASR-Nano | LLM-based 모델입니다. decoder throughput이 중요하면 vLLM을 사용합니다. | [vLLM guide](./vllm_guide.md) |
+| LLM-based ASR 또는 중영일 + 중국어 방언/지역 억양 평가 | Fun-ASR-Nano | Python에서 평가한 뒤 checkpoint와 interface에 맞는 vLLM 경로를 선택합니다. | [vLLM 경로 선택](#vllm-checkpoint-paths) |
 | 오프라인 장시간 ASR 및 익명 화자 라벨 | MOSS-Transcribe-Diarize | 한 번의 오프라인 request로 전사, timestamps, 녹음 내 익명 화자 라벨을 반환합니다. 알려진 인물을 식별하지 않으며 외부 VAD / speaker model도 필요하지 않습니다. | [MOSS deployment guide](./moss_transcribe_diarize.md) |
 | 라이브 자막 또는 콜센터 스트림 | Runtime WebSocket service | 장시간 연결, 부분 결과, endpointing에 맞춘 런타임입니다. | [Runtime service docs](../runtime/readme.md) |
 | Whisper / cloud ASR에서 전환 | SenseVoice-Small로 baseline을 만들고 필요하면 비교 | 강한 기본 경로로 먼저 평가한 뒤 용도별로 조정하는 편이 안전합니다. | [Migration guide](./migration_from_whisper.md) |
@@ -42,6 +42,12 @@ result = model.generate(input="meeting.wav")
 - **`paraformer`**: `paraformer-zh`에 VAD와 문장부호 복원을 결합한 중국어 경로입니다.
 - **`paraformer-en`**: `paraformer-en`과 VAD를 사용하는 OpenAI-style client용 영어 전사입니다.
 - **`fun-asr-nano`**: `FunAudioLLM/Fun-ASR-Nano-2512`로 중영일·중국어 방언/지역 억양을 평가합니다. vLLM acceleration을 시험할 때는 호환되는 runtime을 선택하세요.
+- **`moss-transcribe-diarize`**: 서드파티 `OpenMOSS-Team/MOSS-Transcribe-Diarize`의 오프라인 전사와 녹음 내 익명 화자 라벨입니다. [MOSS guide(영문)](./moss_transcribe_diarize.md)에서 전용 의존성과 remote code를 검토하고, 구조화된 segment에는 `verbose_json`을 요청하세요. 외부 VAD / speaker model이 필요 없으며 알려진 인물을 식별하지 않습니다.
+
+여기서 설명하는 alias는 `AutoModel`을 로드하는 [예제 server](../examples/openai_api/server.py)의 설정입니다.
+native vLLM이나 `AutoModelVLLM`을 자동으로 선택하지 않습니다.
+패키지의 `funasr-server`는 별도 loader / backend 선택 로직을 사용하므로,
+서비스 사이에서 alias나 성능 결과를 그대로 재사용하지 마세요.
 
 이 HTTP 예제는 최상위 `text`와 `verbose_json`의 각 segment `text`를 정리하므로,
 형식을 바꿔도 감정/이벤트 태그가 복원되지 않습니다. 원래 태그가 필요하면 Python SDK를
@@ -57,6 +63,21 @@ python examples/openai_api/smoke_test.py --base-url http://localhost:8000 --mode
 
 SDK, JavaScript, workflow, Postman, OpenAPI, Docker, Kubernetes는 [OpenAI API example](../examples/openai_api/README_ko.md)에서 시작하세요.
 
+<a id="vllm-checkpoint-paths"></a>
+
+## vLLM checkpoint와 interface 선택
+
+| 경로 | checkpoint와 interface | 다음 문서 |
+| --- | --- | --- |
+| FunASR split-engine | 기본 `FunAudioLLM/Fun-ASR-Nano-2512`를 `AutoModelVLLM`으로 로드합니다. FunASR은 오디오 부분, vLLM은 decoder를 처리합니다. | [Split-engine(영문)](./vllm_guide.md) |
+| 공식 native vLLM | 변환된 `FunAudioLLM/Fun-ASR-Nano-2512-vllm`을 vLLM의 native 구현으로 로드하고 `/v1/audio/transcriptions`를 사용합니다. `AutoModelVLLM` 로드가 아닙니다. | [공식 기능 검증(영문)](./vllm_official_native_validation.md) |
+| 과거 community native vLLM | `allendou/Fun-ASR-Nano-2512-vllm`, 2026-08-13 검증입니다. 측정 시간은 당시 checkpoint와 환경에만 해당합니다. | [과거 community 기록](./vllm_native_funasr_validation.md) |
+
+공식 기록은 고정 revision과 기존 환경에서의 기능 검증이며, 신규 설치 절차,
+지속 부하 benchmark 또는 `/v1/realtime` streaming 검증이 아닙니다.
+과거 community 측정 시간을 공식 모델의 결과로 사용하지 마세요.
+MOSS는 별도 가이드를 따르세요. Nano checkpoint와 검증은 MOSS 호환성을 입증하지 않습니다.
+
 ## 벤치마크 후 결정하기
 
 깨끗한 demo 오디오 하나만 보고 모델을 정하지 마세요. 먼저 작은 대표 세트로 확인합니다.