cstr/funasr-nano-GGUF

Fun-ASR-Nano-2512 — GGUF (ggml-quantised)

GGUF / ggml conversion of FunAudioLLM/Fun-ASR-Nano-2512 for use with the funasr backend in CrispStrobe/CrispASR.

Fun-ASR-Nano-2512 is Alibaba's speech-LLM ASR model targeting Mandarin + Cantonese + English + Japanese + Korean:

• 70-block SenseVoiceSmall SANM encoder (1 entry block @ 560→512 + 49 main blocks + 20 "tp" blocks, all 512-dim, 4 heads, FSMN k=11 depthwise convolution branch)
• 2-block Transformer audio adaptor (512 → 2048 → 1024 prelude + 2× MHA blocks at 1024, FFN inner = 256)
• Qwen3-0.6B LLM decoder (28 layers, GQA 16/8, head_dim 128, RoPE θ=1e6, RMSNorm eps=1e-6) — the same body as Qwen3-ASR's decoder
• Speech is spliced into the LLM via the ChatML prompt <|im_start|>user 语音转写：<placeholders><|im_end|>\n<|im_start|>assistant\n and decoded autoregressively
• KV cache so per-token decode is O(1) in cache size

Architecture note — no CTC path

Upstream config.yaml and funasr/models/fun_asr_nano/model.py declare a CTC decoder + head, but the published model.pt ships only audio_encoder.* + audio_adaptor.* + llm.* (1261 tensors total, zero ctc_decoder.* / ctc.ctc_lo.* keys). The LLM-decoder path is therefore the only viable inference path for these weights, and is what this GGUF and the CrispASR runtime implement.

Files

File	Size	Notes
funasr-nano-2512.gguf (alias)	1.98 GB	symlink/alias of the F16
funasr-nano-2512-f16.gguf	1.98 GB	F16, full precision reference
funasr-nano-2512-q8_0.gguf	1.27 GB	Q8_0, near-lossless
funasr-nano-2512-q4_k.gguf	897 MB	Q4_K — recommended default

All three precisions produce byte-identical output on samples/jfk.wav:

AND SO MY FELLOW AMERICANS ASK NOT WHAT YOUR COUNTRY CAN DO FOR YOU ASK WHAT YOU CAN DO FOR YOUR COUNTRY

(Fun-ASR-Nano outputs upper-case English without punctuation; pipe through --punc-model fullstop-punc or fireredpunc if you need proper casing/punctuation.)

Quick Start

git clone https://github.com/CrispStrobe/CrispASR
cd CrispASR
cmake -B build-ninja-compile -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build build-ninja-compile --target crispasr

# Auto-download (recommended Q4_K)
./build-ninja-compile/bin/crispasr -m funasr --auto-download -f samples/jfk.wav

# Or pin a specific file
hf download cstr/funasr-nano-GGUF funasr-nano-2512-q4_k.gguf --local-dir .
./build-ninja-compile/bin/crispasr -m funasr-nano-2512-q4_k.gguf -f samples/jfk.wav

Licence + attribution

Upstream FunAudioLLM/Fun-ASR-Nano-2512:

• Code (the funasr Python package): Apache-2.0.
• Model weights: FunASR Model License v1.1 (Alibaba) — commercial use OK with attribution. Confirmed on the upstream-tracking discussion in CrispStrobe/CrispASR#99.

These GGUF files are a quantised / repackaged distribution of the upstream weights and inherit the FunASR Model License v1.1. Please attribute Alibaba / FunAudioLLM in downstream products.

If you use this model, please also cite the upstream FunASR work. See the upstream model card for the canonical citation.