FunAudioLLM/Fun-ASR-Nano-GGUF

1---2license: apache-2.03language:4- zh5- en6library_name: gguf7tags:8- automatic-speech-recognition9- asr10- fun-asr11- funasr12- qwen313- llama.cpp14- ggml15- cpu16- chinese17- arxiv:2407.0405118pipeline_tag: automatic-speech-recognition19---20 21# Fun-ASR-Nano · GGUF (FunASR llama.cpp runtime)22 23GGUF build of **Fun-ASR-Nano** (SenseVoice SAN-M encoder + adaptor + **Qwen3-0.6B** LLM decoder) for the zero-Python, CPU/edge **[FunASR llama.cpp runtime](https://github.com/FunAudioLLM/Fun-ASR/tree/main/runtime/llama.cpp)** — the accuracy leader (LLM decoder), single C++ binary.24 25## LLM quantization (pick by size vs accuracy)26 27The Fun-ASR-Nano LLM (Qwen3-0.6B) ships in three tiers — all within 0.1% CER (184-file micro-CER). Pair any with `funasr-encoder-f16.gguf` (470 MB).28 29| LLM file | size | CER ↓ | speed |30|---|---|---|---|31| `qwen3-0.6b-q4km.gguf` | **484 MB** | 8.35% | 6.1× | smallest |32| `qwen3-0.6b-q5km.gguf` | 551 MB | **8.25%** | 5.7× | best accuracy |33| `qwen3-0.6b-q8_0.gguf` | 805 MB | 8.30% | 6.0× | |34 35Recommended: **q4_K_M** (smallest) or **q5_K_M** (best).36 37## Get it running (no Python, no build)38 39These are GGUF weights for the **[FunASR llama.cpp runtime](https://github.com/modelscope/FunASR/tree/main/runtime/llama.cpp)** — a whisper.cpp-style, single self-contained binary for CPU / edge. Grab a prebuilt binary, then fetch this model and run:40 41- **Prebuilt binaries (Linux / macOS / Windows) → [GitHub Releases](https://github.com/modelscope/FunASR/releases)** (tag `runtime-llamacpp-v*`)42- **Deployment guide & qualified benchmarks → [funasr.com/deploy/llama-cpp](https://www.funasr.com/deploy/llama-cpp.html)**43 44```bash45bash download-funasr-model.sh nano ./gguf46llama-funasr-cli --enc ./gguf/funasr-encoder-f16.gguf -m ./gguf/qwen3-0.6b-q8_0.gguf --vad ./gguf/fsmn-vad.gguf -a audio.wav47```48 49## Files50| file | size | notes |51|---|---|---|52| `funasr-encoder-f16.gguf` | 470 MB | audio encoder + adaptor (f16) |53| `qwen3-0.6b-q8_0.gguf` | 805 MB | LLM decoder, **recommended** (Q8_0) |54| `qwen3-0.6b-q4km.gguf` | 484 MB | LLM decoder, smaller (Q4_K_M) |55 56## Usage (needs both the encoder and the LLM gguf)57```bash58llama-funasr-cli --enc funasr-encoder-f16.gguf -m qwen3-0.6b-q8_0.gguf -a audio.wav --vad fsmn-vad.gguf59```60On CPU: **8.30 % CER** on the 184-clip Mandarin benchmark (vs whisper.cpp 22–31 %).61 62## Links63- 🧩 Runtime & build: **[Fun-ASR · runtime/llama.cpp](https://github.com/FunAudioLLM/Fun-ASR/tree/main/runtime/llama.cpp)** — ⭐ **Star [Fun-ASR](https://github.com/FunAudioLLM/Fun-ASR)!**64- Source model: [FunAudioLLM/Fun-ASR-Nano-2512](https://huggingface.co/FunAudioLLM/Fun-ASR-Nano-2512)65