FunAudioLLM/Fun-ASR-Nano-2512-GGUF

Fun-ASR-Nano-2512 GGUF

Standalone audio.cpp GGUF builds of FunAudioLLM/Fun-ASR-Nano-2512-hf. Each file embeds the model configuration, processor configuration, tokenizer, chat template, and the audio.cpp model package specification.

Files

The source checkpoint is pinned to revision 854d88f94205cd17d2afdb24332130d86fbe654a. The source model.safetensors SHA256 is 335ca3e74917f1156690400e2c344350112950165789cf78ce3d0a367affd821.

audio.cpp

audiocpp_cli \
  --task asr \
  --family fun_asr_nano \
  --model fun-asr-nano-2512-q8_0.gguf \
  --backend cuda \
  --audio speech.wav

Fun-ASR-Nano currently provides offline multilingual ASR. It does not expose streaming or timestamp output. On CUDA, audio.cpp keeps the Q80 encoder and adaptor weights native and loads decoder weights as BF16 by default for stable logits. An explicit `funasrnano.decoderweight_type` session option overrides that default.

Reproducibility

The files were generated with audio.cpp's audiocpp_gguf converter:

audiocpp_gguf \
  --input model.safetensors \
  --root /path/to/Fun-ASR-Nano-2512-hf \
  --output fun-asr-nano-2512-q8_0.gguf \
  --type q8_0 \
  --family fun_asr_nano \
  --model-spec model_specs/fun_asr_nano.json

Both formats were checked with audiocpp_gguf --inspect and full reference audio transcription on CPU and NVIDIA H100 CUDA.

License

The original model and these converted weights are governed by the FunASR Model Open Source License Agreement v1.1 distributed with the source model. Review that agreement before using or redistributing the files.