modelscope · LauraGPT · Sep 9, 2026 · Sep 9, 2026
diff --git a/README.md b/README.md
@@ -357,7 +357,7 @@ recordings with the same evaluation scope.
 
 - **MOSS-Transcribe-Diarize** brings long-form ASR, timestamps, and anonymous speaker labels to FunASR services, Docker, Kubernetes, vLLM/SGLang workflows, and FunClip. [Deploy MOSS ->](./docs/moss_transcribe_diarize.md)
 - **FunASR 1.4.15** adds tested NumPy 2 compatibility and fixes streaming KWS/VAD boundaries and checkpoint ranking. Install with `python -m pip install -U "funasr==1.4.15"`. [Release and verification scope ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.15)
-- **Production deployment** now includes faster, more resilient realtime serving and verified llama.cpp archives for ten Linux, macOS, and Windows targets. [GPU services ->](./docs/vllm_guide.md) · [CPU/edge packages ->](https://www.funasr.com/en/deploy/llama-cpp.html)
+- **Native Transformers:** Fun-ASR-Nano [merged upstream](https://github.com/huggingface/transformers/pull/46180). Use the official `-hf` checkpoint and pinned source; stable 5.16.1 does not include it. [Installation and inference ->](./docs/transformers_native.md)
 
 > See [GitHub Releases](https://github.com/modelscope/FunASR/releases) for the complete changelog and downloadable assets.
 

diff --git a/README_ja.md b/README_ja.md
@@ -105,7 +105,7 @@ CER/WER をそろえて比較してください。オフラインのスループ
 
 - **MOSS-Transcribe-Diarize** を FunASR service、Docker、Kubernetes、vLLM/SGLang workflow、FunClip に統合し、長時間 ASR、timestamp、匿名 speaker label を一度に処理できます。[MOSS をデプロイ ->](./docs/moss_transcribe_diarize.md)
 - **FunASR 1.4.15** はテスト済みの NumPy 2 互換性を追加し、ストリーミング KWS/VAD の境界処理と checkpoint の順位付けを修正します。`python -m pip install -U "funasr==1.4.15"`。[リリースと検証範囲 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.15)
-- **Production deployment** に、より高速で安定した realtime serving と、Linux、macOS、Windows の 10 target 向け llama.cpp package を追加しました。[GPU service ->](./docs/vllm_guide.md) · [CPU / edge package ->](https://www.funasr.com/en/deploy/llama-cpp.html) · [v0.2.6 binaries ->](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.6)
+- **ネイティブ Transformers：** Fun-ASR-Nano が[上流にマージ](https://github.com/huggingface/transformers/pull/46180)されました。公式 `-hf` checkpoint と固定ソースを使用します。安定版 5.16.1 には未収録です。[導入ガイド（英語） ->](./docs/transformers_native.md)
 
 > 完全な変更履歴と download asset は [GitHub Releases](https://github.com/modelscope/FunASR/releases) を参照してください。
 

diff --git a/README_ko.md b/README_ko.md
@@ -104,7 +104,7 @@ checkpoint/revision, 오디오 집합, 하드웨어, 배치, 워밍업, 측정
 
 - **MOSS-Transcribe-Diarize**를 FunASR service, Docker, Kubernetes, vLLM/SGLang workflow, FunClip에 통합해 긴 오디오 ASR, timestamp, 익명 speaker label을 한 번에 처리합니다. [MOSS 배포 ->](./docs/moss_transcribe_diarize.md)
 - **FunASR 1.4.15**는 테스트를 거친 NumPy 2 호환성을 추가하고 스트리밍 KWS/VAD 경계 처리와 checkpoint 순위 산정을 수정합니다. `python -m pip install -U "funasr==1.4.15"`. [릴리스 및 검증 범위 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.15)
-- **Production deployment**에 더 빠르고 안정적인 realtime serving과 Linux, macOS, Windows 10개 target용 llama.cpp package를 추가했습니다. [GPU service ->](./docs/vllm_guide.md) · [CPU / edge package ->](https://www.funasr.com/en/deploy/llama-cpp.html) · [v0.2.6 binaries ->](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.6)
+- **네이티브 Transformers:** Fun-ASR-Nano가 [업스트림에 병합](https://github.com/huggingface/transformers/pull/46180)되었습니다. 공식 `-hf` checkpoint와 고정 소스를 사용하세요. 안정 버전 5.16.1에는 아직 포함되지 않았습니다. [설치 가이드(영문) ->](./docs/transformers_native.md)
 
 > 전체 변경 기록과 download asset은 [GitHub Releases](https://github.com/modelscope/FunASR/releases)에서 확인할 수 있습니다.
 

diff --git a/README_zh.md b/README_zh.md
@@ -150,7 +150,7 @@ checkpoint/revision、音频集、硬件、批量大小、预热、计时范围
 
 - **MOSS-Transcribe-Diarize** 已接入 FunASR 服务、Docker、Kubernetes、vLLM/SGLang 工作流和 FunClip，一次完成长音频转写、时间戳与匿名说话人标注。[部署 MOSS ->](./docs/moss_transcribe_diarize_zh.md)
 - **FunASR 1.4.15** 新增经过测试的 NumPy 2 兼容支持，修复流式 KWS/VAD 边界处理和 checkpoint 排序。升级命令：`python -m pip install -U "funasr==1.4.15"`。[发布说明与验证范围 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.15)
-- **工业部署** 新增更快、更稳定的实时服务，以及覆盖 Linux、macOS、Windows 十种目标的 llama.cpp 预编译包。[GPU 服务 ->](./docs/vllm_guide_zh.md) · [CPU/端侧包 ->](https://www.funasr.com/deploy/llama-cpp.html)
+- **原生 Transformers：** Fun-ASR-Nano [已合入上游](https://github.com/huggingface/transformers/pull/46180)。使用官方 `-hf` checkpoint 与固定源码；稳定版 5.16.1 尚未包含。[安装与推理 ->](./docs/transformers_native_zh.md)
 
 > 完整改动记录和可下载资产请查看 [GitHub Releases](https://github.com/modelscope/FunASR/releases)。
 

diff --git a/docs/README.md b/docs/README.md
@@ -12,6 +12,7 @@
 | Install and check the Python environment | [Installation](installation/installation.md) / [中文](installation/installation_zh.md) |
 | Choose a model, language, or timestamp capability | [Model selection](model_selection.md) / [中文](model_selection_zh.md) |
 | Run the first transcription | [Quickstart](tutorial/README.md) / [中文](tutorial/README_zh.md) / [CLI](cli.md) |
+| Use native Hugging Face Transformers APIs | [Fun-ASR-Nano native guide](transformers_native.md) / [中文](transformers_native_zh.md) |
 | Configure AutoModel, generate, and streaming cache | [Python SDK](python_api.md) / [中文](python_api_zh.md) |
 | Move from Whisper or a cloud API | [Migration](migration_from_whisper.md) / [中文](migration_from_whisper_zh.md) |
 

diff --git a/docs/deployment_matrix.md b/docs/deployment_matrix.md
@@ -6,6 +6,7 @@ Use this page to choose the shortest deployment path for a product, demo, benchm
 
 | Path | Best for | Start here | Operational notes |
 |---|---|---|---|
+| Native Transformers | Python applications using Hugging Face processors and generation | [Native Fun-ASR-Nano guide](./transformers_native.md) | Official `-hf` checkpoint and pinned source; stable 5.16.1 lacks support at the 2026-09-09 check. Python inference, not an HTTP or realtime server. |
 | Colab notebook | Browser smoke tests, first evaluation, shareable demos | [Colab quickstart](../examples/colab/) | No local setup; first run downloads model files, GPU runtime is faster. |
 | Python API | Notebooks, offline jobs, first model evaluation | [README quick start](../README.md#quick-start) | Lowest ceremony; caller owns batching, retries, and files. |
 | OpenAI-compatible API | Private speech API, agents, Dify/LangChain/AutoGen-style clients | [OpenAI API example](../examples/openai_api/) | Easiest integration for apps that already support OpenAI audio APIs. |

diff --git a/docs/deployment_matrix_zh.md b/docs/deployment_matrix_zh.md
@@ -6,6 +6,7 @@
 
 | 路径 | 适合场景 | 从这里开始 | 运维提示 |
 |---|---|---|---|
+| 原生 Transformers | 已使用 Hugging Face 处理器与生成接口的 Python 应用 | [原生 Fun-ASR-Nano 指南](./transformers_native_zh.md) | 官方 `-hf` checkpoint 与固定源码；2026-09-09 核验的稳定版 5.16.1 尚未包含。是 Python 推理入口，不是 HTTP 或实时服务器。 |
 | Colab Notebook | 浏览器 smoke test、首次评估、可分享 demo | [Colab 快速体验](../examples/colab/README_zh.md) | 不需要本地环境；首次运行会下载模型，GPU runtime 更快。 |
 | Python API | Notebook、离线任务、首次模型评测 | [README 快速开始](../README_zh.md#快速开始) | 最简单；调用方自己负责批处理、重试和文件管理。 |
 | OpenAI 兼容 API | 私有语音 API、Agent、Dify/LangChain/AutoGen 风格客户端 | [OpenAI API 示例](../examples/openai_api/README_zh.md) | 已支持 OpenAI audio API 的应用最容易接入。 |

diff --git a/docs/transformers_native.md b/docs/transformers_native.md
@@ -0,0 +1,208 @@
+# Fun-ASR-Nano with native Transformers
+
+[简体中文](./transformers_native_zh.md) | English
+
+Use this path when your Python application already works with Hugging Face
+processors and generation APIs. It runs Fun-ASR-Nano through native Transformers
+classes, without the FunASR toolkit or checkpoint-provided remote Python code.
+For a managed HTTP service, start with the [deployment matrix](./deployment_matrix.md);
+this guide does not add an OpenAI endpoint, request queue or realtime server.
+
+## What was merged, and what was released?
+
+On **2026-09-09**, [Transformers PR #46180](https://github.com/huggingface/transformers/pull/46180)
+was merged as `fc501343edfccdc840eb8594a6cafa8185c2de53`.
+At verification time that day, the latest stable **5.16.1 did not contain this
+model**: both its release tag and actual wheel were inspected. A merge is not
+a stable release. The source build identifies itself as `5.17.0.dev0`; that is
+not a promised release version or date. Pin the tested source below, or check
+that a later released package really contains `fun_asr_nano` before migrating.
+
+## Choose the checkpoint and the interface together
+
+| Path | Checkpoint | Entry point |
+| --- | --- | --- |
+| FunASR toolkit | `FunAudioLLM/Fun-ASR-Nano-2512` | `funasr.AutoModel`; separate documented split-engine path |
+| Native Transformers | `FunAudioLLM/Fun-ASR-Nano-2512-hf` | `AutoProcessor` + `AutoModelForSpeechSeq2Seq` |
+| Native vLLM | `FunAudioLLM/Fun-ASR-Nano-2512-vllm` | [Official native vLLM guide](./vllm_official_native_validation.md) |
+| Native C++ | Converted GGUF files for the chosen runtime | [llama.cpp guide](../runtime/llama.cpp/README.md) |
+
+These are different artifacts and API contracts. Renaming a model ID or passing
+`backend="vllm"` does not convert weights. Native Transformers support does not
+make the `-hf` checkpoint a vLLM, GGUF or FunASR HTTP-server checkpoint.
+
+The official [native HF checkpoint](https://huggingface.co/FunAudioLLM/Fun-ASR-Nano-2512-hf/tree/d93b302ee7fd505e1b3576120fc142fc6f7820e1)
+is public and ungated at revision `d93b302ee7fd505e1b3576120fc142fc6f7820e1`.
+Its configuration maps to `FunAsrNanoForConditionalGeneration`. Its feature
+extractor is embedded in `processor_config.json`; a separate
+`preprocessor_config.json` is not required for this snapshot.
+
+## Independent CPU environment
+
+Use a new directory for the virtual environment; do not upgrade an existing
+FunASR or vLLM service environment in place. The following is **Linux x86-64,
+Python 3.12, CPU-only**, not a CUDA or macOS installation recipe.
+
+```bash
+python3.12 -m venv .venv-funasr-native
+. .venv-funasr-native/bin/activate
+python -m pip install --index-url https://download.pytorch.org/whl/cpu \
+  'torch==2.10.0+cpu' 'torchaudio==2.10.0+cpu'
+python -m pip install 'numpy==1.26.4' 'librosa==0.11.0' 'soundfile==0.13.1' \
+  'huggingface-hub==1.30.0' 'tokenizers==0.23.2' \
+  'https://github.com/huggingface/transformers/archive/fc501343edfccdc840eb8594a6cafa8185c2de53.zip'
+python -m pip check
+```
+
+The verification environment installed a wheel built from the exact source
+commit above. Matching **torch and torchaudio** are required for this native
+feature extractor: it calls `torchaudio.compliance.kaldi.fbank`. The toolkit's
+optional audio dependencies do not imply that torchaudio is optional here.
+Older prepared environments failed dependency gates; do not disable those gates
+or mix incompatible Hub/tokenizers versions to get an import to succeed.
+This is a tested environment, not a complete lock of every transitive dependency;
+record `python -m pip freeze`, platform details and model revision in your own run.
+
+## Check preprocessing without loading weights
+
+This downloads the public configuration, tokenizer and chat template, **not
+model weights**. One second of synthetic silence checks the entry point; it is
+not a speech-recognition quality test.
+
+<!-- native-example: processor -->
+```python
+import numpy as np
+from transformers import AutoProcessor
+
+model_id = "FunAudioLLM/Fun-ASR-Nano-2512-hf"
+revision = "d93b302ee7fd505e1b3576120fc142fc6f7820e1"
+processor = AutoProcessor.from_pretrained(
+    model_id, revision=revision, trust_remote_code=False, token=False
+)
+inputs = processor.apply_transcription_request(
+    audio=np.zeros(16000, dtype=np.float32),
+    language="en",
+    processor_kwargs={
+        "return_tensors": "pt",
+        "audio_kwargs": {"sampling_rate": 16000},
+    },
+)
+print({name: tuple(value.shape) for name, value in inputs.items()})
+```
+
+In the verified environment the text tensors have shape `(1, 42)`, audio features
+`(1, 17, 560)`, and feature mask `(1, 17)`. There are 17 audio placeholder tokens
+and 17 valid feature frames. These shapes depend on input length and template;
+do not hard-code them for real recordings.
+
+## Transcribe one recording
+
+Prepare a non-empty **mono 16 kHz WAV** named `audio.wav` in your working directory.
+The code deliberately rejects a different sample rate or channel layout rather
+than silently assigning 16 kHz to arbitrary samples. Resample/downmix with an
+explicit policy before this step and preserve the original recording for review.
+
+The first model load downloads the weights unless the pinned snapshot is already
+cached. It needs substantially more memory and time than preprocessing. This
+example uses CPU float32 explicitly; GPU placement, mixed precision, attention
+backends and service concurrency need separate validation.
+
+<!-- native-example: transcribe -->
+```python
+import soundfile as sf
+import torch
+from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
+
+model_id = "FunAudioLLM/Fun-ASR-Nano-2512-hf"
+revision = "d93b302ee7fd505e1b3576120fc142fc6f7820e1"
+audio, sample_rate = sf.read("audio.wav", dtype="float32")
+if sample_rate != 16000 or audio.ndim != 1 or audio.size == 0:
+    raise ValueError("Use a non-empty mono 16 kHz WAV file")
+
+processor = AutoProcessor.from_pretrained(
+    model_id, revision=revision, trust_remote_code=False, token=False
+)
+model = AutoModelForSpeechSeq2Seq.from_pretrained(
+    model_id, revision=revision, trust_remote_code=False, token=False,
+    dtype=torch.float32,
+).to("cpu").eval()
+inputs = processor.apply_transcription_request(
+    audio=audio, language="zh",
+    processor_kwargs={
+        "return_tensors": "pt",
+        "audio_kwargs": {"sampling_rate": sample_rate},
+    },
+)
+with torch.inference_mode():
+    generated = model.generate(**inputs, max_new_tokens=128, do_sample=False)
+new_tokens = generated[:, inputs.input_ids.shape[1]:]
+print(processor.batch_decode(new_tokens, skip_special_tokens=True)[0])
+```
+
+`generated` includes the input prompt. Slice off `inputs.input_ids.shape[1]`
+before decoding so prompt framing is not mistaken for recognized speech.
+`max_new_tokens=128` bounds this short-file example; hitting the limit can
+truncate output. A longer limit is not proof of complete coverage.
+Use `language="en"` or `language="ja"` for the corresponding target language;
+language selection is a transcription instruction, not a promise of translation.
+
+## Context, hotwords and batches
+
+`apply_transcription_request` accepts `prompt` for context and `keywords` for
+hotwords. These are native processor arguments, not the toolkit's `hotword`
+parameter or a vLLM HTTP field. The source API accepts one language for the batch
+or a language list matching its length; per-recording prompt and nested keyword
+lists must also match the batch. Keep the decoded results in input order.
+
+Do not infer quality gains from a template accepting a keyword. Evaluate names,
+numbers and unusual terms against authorized, representative recordings.
+Batch padding, memory use and generated-text correctness must be tested together
+before turning a single-file example into a production batch worker.
+
+## Evidence and limits
+
+The documented single-recording recipe was also executed on the official Chinese
+sample, followed by an English single recording, a Chinese keyword request and a
+Chinese/English batch. All returned non-empty text and EOS before the 128-token
+limit. These are functional checks on two short public samples, not a CER/WER or
+capacity evaluation. CPU: Intel Xeon Platinum 8480+, float32, four Torch threads;
+matching torch/torchaudio 2.10.0+cpu and librosa 0.11.0. Model weights were loaded
+from a previously downloaded official cache whose SHA/size were rechecked.
+
+Samples came from the original model's
+[fixed example directory](https://huggingface.co/FunAudioLLM/Fun-ASR-Nano-2512/tree/272c57b82523ada6fd87095e955f8e29100979ab/example),
+not from an invented -hf example directory. The mono 48 kHz Chinese MP3 was
+explicitly resampled with librosa's default soxr_hq and written as float32 16 kHz
+WAV before running the unmodified recipe. Its raw text was
+`开饭时间早上九点至下午五点。`; adding the candidate keyword `开放时间` did not force that
+spelling into the output. A keyword is a hint, not an enforced vocabulary.
+
+Synthetic processor tests exercised language aliases, context/keyword lists and
+two-length batches. The empty-list case still raises an upstream `IndexError`
+before its intended validation; reject empty batches in your application.
+The guide's explicit non-empty waveform check avoids that entry path. Successful
+cases do not erase this observed limitation. Initial resampling also exposed the
+old librosa 0.10.1 dependency on removed `pkg_resources`; the final environment
+uses 0.11.0 and was retested.
+
+The 2026-09-09 isolated CPU verification passed package dependency checks,
+native configuration/class resolution and synthetic preprocessing with the
+pinned official revision. Installed configuration, modeling, processing and
+feature-extraction files were byte-identical to the merged upstream source.
+A preprocessing pass is **not an accuracy or capacity result**. It does not
+establish word timestamps, diarization, realtime behavior, GPU compatibility,
+or performance equivalence with vLLM.
+
+For anonymous speakers and segment timestamps, see the third-party OpenMOSS
+[MOSS guide](./moss_transcribe_diarize.md); for other choices see the
+[Model Zoo](../model_zoo/readme.md). Do not infer timestamp or identity support
+from a model returning text. Preserve input hashes, versions, raw responses and
+an authorized human review when assessing an application. Never attach private
+recordings or credentials to a public issue.
+
+## Sources and next steps
+
+- [Merged implementation and usage documentation](https://github.com/huggingface/transformers/blob/fc501343edfccdc840eb8594a6cafa8185c2de53/docs/source/en/model_doc/fun_asr_nano.md).
+- [Native processor](https://github.com/huggingface/transformers/blob/fc501343edfccdc840eb8594a6cafa8185c2de53/src/transformers/models/fun_asr_nano/processing_fun_asr_nano.py) and [audio feature extractor](https://github.com/huggingface/transformers/blob/fc501343edfccdc840eb8594a6cafa8185c2de53/src/transformers/models/fun_asr_nano/feature_extraction_fun_asr_nano.py).
+- [Why the checkpoint suffix matters](https://www.funasr.com/en/blog/fun-asr-nano-transformers.html), an application-oriented walkthrough.
+- [Fun-ASR model project](https://github.com/QwenAudio/Fun-ASR) and [FunASR toolkit](https://github.com/modelscope/FunASR).