ALPHA — EXPERIMENTAL The inference engine (s2.cpp) is an early-stage, community-built project. Expect rough edges and breaking changes. Not production-ready.

GGUF-quantized weights of Fish Audio S2 Pro, a high-quality multilingual text-to-speech model with voice cloning support, packaged for local inference with s2.cpp — a pure C++/GGML engine with no Python dependency.

License: Fish Audio Research License — free for research and non-commercial use. Commercial use requires a separate license from Fish Audio. See LICENSE.md and fish.audio.

Files

File	Size
`s2-pro-f16.gguf`	9.3 GB
`s2-pro-q8_0.gguf`	5.3 GB
`s2-pro-q6_k.gguf`	4.3 GB
`s2-pro-q5_k_m.gguf`	3.8 GB
`s2-pro-q4_k_m.gguf`	3.4 GB
`s2-pro-q3_k.gguf`	2.9 GB
`s2-pro-q2_k.gguf`	2.4 GB
`tokenizer.json`	12 MB

All GGUF files contain both the transformer weights and the audio codec in a single file.

Requirements

GPU with Vulkan support (AMD/NVIDIA/Intel) or CPU with enough RAM
s2.cpp built from source (C++17 + CMake)

VRAM guide

VRAM	Recommended
≥ 8 GB	`q8_0`
6–8 GB	`q6_k`
4–6 GB	`q5_k_m`
3–4 GB	`q4_k_m`
< 3 GB	`q3_k` / `q2_k` (quality degrades)
CPU only	`q4_k_m` or lower (slow)

Quick start

bash

# Clone and build s2.cpp
git clone --recurse-submodules https://github.com/rodrigomatta/s2.cpp.git
cd s2.cpp
cmake -B build -DCMAKE_BUILD_TYPE=Release -DS2_VULKAN=ON
cmake --build build --parallel $(nproc)

# Download model files (example with huggingface-cli)
huggingface-cli download rodrigomt/s2-pro-gguf s2-pro-q6_k.gguf tokenizer.json --local-dir .

# Synthesize
./build/s2 \
  -m s2-pro-q6_k.gguf \
  -t tokenizer.json \
  -text "Hello, this is a test." \
  -v 0 \
  -o output.wav

Voice cloning

bash

./build/s2 \
  -m s2-pro-q6_k.gguf \
  -t tokenizer.json \
  -pa reference.wav \
  -pt "Transcript of the reference audio." \
  -text "Text to synthesize in that voice." \
  -v 0 \
  -o output.wav

Reference audio: 5–30 seconds, clean recording, WAV or MP3.

Quantization

The GGUF files in this repository were re-quantized using a modified llama-quantize tool with improved quantization routines.

Model architecture

S2 Pro uses a Dual-AR architecture (~4.56B parameters total):

Slow-AR — 36-layer Qwen3 transformer (4.13B params), GQA (32 heads / 8 KV heads), RoPE 1M base, persistent KV cache
Fast-AR — 4-layer transformer (0.42B params) generating 10 acoustic codebook tokens per semantic step
Audio codec — convolutional RVQ encoder/decoder (10 codebooks × 4096 entries)

License

The model weights are licensed under the Fish Audio Research License.

Research and non-commercial use: free under this license
Commercial use: requires a separate written license from Fish Audio

Full terms: LICENSE.md · Commercial: fish.audio · business@fish.audio