s2-pro-GGUF

Fish Audio S2 Pro is a state-of-the-art multilingual text-to-speech (TTS) model from Fish Audio that combines a dual-autoregressive (Dual-AR) transformer architecture with an RVQ-based audio codec (10 codebooks, ~21 Hz) — a 4B-parameter "Slow AR" module predicts the primary semantic codebook along the time axis while a 400M-parameter "Fast AR" generates the 9 residual codebooks for fine acoustic detail — and is trained on over 10 million hours of audio across 80+ languages (with Japanese, English, and Chinese as Tier 1). Its standout feature is fine-grained inline control: free-form natural-language instructions like [whisper], [laughing], or [pitch up] can be embedded directly in the text (15,000+ unique tags supported) to steer prosody and emotion at the word level, and it supports multi-speaker, multi-turn generation with low-latency streaming (RTF 0.195 and ~100 ms time-to-first-audio on a single H200) thanks to LLM-native serving optimizations in SGLang such as continuous batching, paged KV cache, and prefix caching. Released with model weights, fine-tuning code, and a streaming inference engine under the Fish Audio Research License, it is free for research and non-commercial use, with commercial use requiring a separate license.

Model Files

File Name Quant Type File Size File Link Description
s2-pro.BF16.gguf BF16 8.07 GB Link Full BF16 weights. Highest quality, largest file size.
s2-pro.Q3_K_L.gguf Q3_K_L 2.25 GB Link Lower quality but usable, good for low RAM availability.
s2-pro.Q3_K_M.gguf Q3_K_M 2.08 GB Link Low quality, smallest file size.
s2-pro.Q4_K_M.gguf Q4_K_M 2.51 GB Link Good quality, default size for most use cases, recommended.
s2-pro.Q4_K_S.gguf Q4_K_S 2.39 GB Link Slightly lower quality with more space savings, recommended.
s2-pro.Q5_K_M.gguf Q5_K_M 2.9 GB Link High quality, recommended.
s2-pro.Q5_K_S.gguf Q5_K_S 2.83 GB Link High quality with slight space savings, recommended.
s2-pro.Q6_K.gguf Q6_K 3.31 GB Link Very high quality, near perfect, recommended.
s2-pro.Q8_0.gguf Q8_0 4.29 GB Link Extremely high quality, generally unneeded but max available quant.
s2-pro.mmproj-bf16.gguf mmproj-bf16 2.05 GB Link Multimodal projection file in BF16 format. Used for audio/speech capabilities.

Run with llama.cpp

Install dependencies and clone llama.cpp

!pip install transformers huggingface-hub

!git clone https://github.com/ggml-org/llama.cpp.git
%cd /data/llama.cpp

Download GGUF model files

Download the BF16 model and BF16 multimodal projector:

from huggingface_hub import snapshot_download

# Target local directory
local_dir = "./model/s2-pro-GGUF"

# Download only the specified files
snapshot_download(
    repo_id="prithivMLmods/s2-pro-GGUF",
    local_dir=local_dir,
    allow_patterns=[
        "s2-pro.BF16.gguf",
        "s2-pro.mmproj-bf16.gguf",
    ],
    resume_download=True,
)

Build llama-tts

Build the TTS CLI with the CUDA backend enabled:

!cmake --build . --config Release -j$(nproc) --target llama-tts

Text-to-speech generation (GPU offloading)

Run inference using llama-tts with GPU offloading enabled:

!/data/llama.cpp/build/bin/llama-tts \
    -m /data/llama.cpp/model/s2-pro-GGUF/s2-pro.BF16.gguf \
    -mm /data/llama.cpp/model/s2-pro-GGUF/s2-pro.mmproj-bf16.gguf \
    -p "Fish Audio S2 Pro is a leading text-to-speech (TTS) model with fine-grained inline control of prosody and emotion. Trained on over 10 million hours of audio data across more than 80 languages, the system combines reinforcement learning alignment with a dual-autoregressive architecture." \
    -ngl 99 \
    -n 1000 \
    --seed 42 \
    --output /data/llama.cpp/model/s2-pro-GGUF/out.wav

Output

The generated audio file will be saved at:

/data/llama.cpp/model/s2-pro-GGUF/out.wav

You can play the generated WAV file using an audio player or a notebook audio widget.

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp

Quantized using llama.cpp by prithivMLmods.

Downloads last month
279
GGUF
Model size
4B params
Architecture
qwen3
Hardware compatibility
Log In to add your hardware

3-bit

4-bit

5-bit

6-bit

8-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for prithivMLmods/s2-pro-GGUF

Base model

fishaudio/s2-pro
Quantized
(17)
this model

Collections including prithivMLmods/s2-pro-GGUF