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

File	Size	SHA256
fun-asr-nano-2512-q8_0.gguf	1,045,334,432 bytes	4d727357574b079b7f43336b2930f39da086ca02f5d8d50872090b4c1c3d5e0a
fun-asr-nano-2512-f16.gguf	1,675,708,832 bytes	3d906c3ccfed07efef88ff53d6cc94b788b9d2edf1492a5679d041b43e98c5be

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 Q8_0 encoder and adaptor weights native and loads decoder weights as BF16 by default for stable logits. An explicit fun_asr_nano.decoder_weight_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.