Model reference · open weights
SenseVoiceSmall-audiocpp is an open-weight audio or speech model from FunAudioLLM. SenseVoiceSmall-GGUF-audiocpp (GGUF) weighs 254 MB; the smallest configuration that runs it is RTX 3060 12 GB.
For self-hosting, compare this build's requirements below with the GPU you plan to rent.
Summary of the FunAudioLLM/SenseVoiceSmall-GGUF-audiocpp model card, 2026-10-01
What it is
| Released by | FunAudioLLM |
|---|---|
| Released | 2026-08-12 |
| VRAM | 254 MB for the weights |
What it runs on
| Card | One stream | Memory |
|---|---|---|
| RTX 3060 12 GB | fits | 11.6 GB |
| RTX 4060 Ti 16 GB | fits | 15.4 GB |
| RTX 3090 24 GB | fits | 23.4 GB |
| RTX 4090 24 GB | fits | 23.4 GB |
| RTX 5090 32 GB | fits | 31.0 GB |
| L40S 48 GB | fits | 44.0 GB |
| A100 80 GB | fits | 78.2 GB |
| H100 80 GB | fits | 78.1 GB |
| RTX PRO 6000 Blackwell 96 GB | fits | 93.8 GB |
| DGX Spark (GB10) 128 GB unified | fits | 107 GB |
| H200 141 GB | fits | 138 GB |
| B200 180 GB | fits | 176 GB |
From the model card
This repository contains a self-contained Q8_0 export of
SenseVoiceSmall for the
audio.cpp spec-backed runtime. The GGUF
embeds the sense_asr schema-v1 model specification, SenseVoice metadata,
SentencePiece vocabulary, CMVN tensors, and 919 model tensors. It loads
directly without --model-spec-override.
| File | Size | SHA256 |
|---|---|---|
sensevoice-small-q8-audiocpp-v1.gguf | 254,211,200 bytes | 4dedf169f625437fb336f2959674f399819729a765e184128c0e25a6e16ff0ec |
audiocpp_cli --task asr --family sense_asr \
--model sensevoice-small-q8-audiocpp-v1.gguf \
--backend cpu --audio zh.wav \
--request-option audio_chunk_mode=none
The integration is tracked in audio.cpp pull request #218.
The model was exported from FunAudioLLM/SenseVoiceSmall revision
3847d57b6bdf2dd8875cb1508d2af43d80a16bf7 with the official
runtime/llama.cpp/export_sensevoice_gguf.py exporter using --wtype q8_0
and --model-spec.
On the official 5.616-second Mandarin sample, direct CPU inference produced:
开饭时间早上9点至下午5点。
The text exactly matched the original Q8 model loaded with an external model specification.
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.
How it works