AIRsLight/Fun-ASR-Nano-2512-GGUF

Fun-ASR-Nano-2512 Q8 GGUF encoder

Q8_0 audio encoder and adaptor export of FunAudioLLM/Fun-ASR-Nano-2512.

This is only the audio encoder half of Fun-ASR-Nano. The FunASR GGUF runtime also requires a compatible Qwen3-0.6B language-model GGUF. The tested pairing uses the official qwen3-0.6b-q4km.gguf from FunAudioLLM/Fun-ASR-Nano-GGUF.

File

File	Size	Quantization
funasr-encoder-q8_0.gguf	241.8 MiB	Q8_0 matrix weights; norms, biases, and FSMN weights remain F32

The encoder was exported directly from the official BF16/FP32 checkpoint with FunASR's upstream exporter. It was not requantized from the F16 GGUF.

Measured behavior

Measurements used the persistent CPU FunASR GGUF runtime on an Intel Core i7-14700KF. Both encoders used the same Q4_K_M language model and decoding configuration.

Encoder	Encoder size	Resident peak working set	Mean recognition time over 59 clips
F16 baseline	447.6 MiB	1337.5 MiB	571 ms
Q8_0	241.8 MiB	1131.5 MiB	590 ms

Q8_0 matched the final F16 transcript exactly on 47 of 59 clips. Aggregate character edit distance from F16 was 2.2%. This is output drift relative to F16, not a labeled character-error-rate measurement. Validate terminology and short noisy speech with your own corpus.

Compatibility

The encoder uses the custom funasr-sensevoice-encoder GGUF architecture in FunASR's runtime/llama.cpp/fun-asr-nano runtime. It cannot be loaded as a standalone language model by general-purpose llama.cpp applications.

Integrity

source model.pt
size:   1971149431 bytes
sha256: 55ae0d2fee369f0f11cce0795f6927934ad17cf11b278a7e56a51272074160bb

funasr-encoder-q8_0.gguf
size:   253553728 bytes
sha256: 5aa6af0e6efad175d66525b87b4531b162733f25449a74667c73c7c3bfd28e10