rodrigomt/s2-pro-gguf

S2 Pro — GGUF

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	Notes
s2-pro-f16.gguf	9.3 GB	Full precision — reference quality
s2-pro-q8_0.gguf	5.7 GB	Near-lossless — recommended for 8+ GB VRAM
s2-pro-q6_k.gguf	4.8 GB	Good quality/size balance — recommended for 6+ GB VRAM
tokenizer.json	—	Qwen3 BPE tokenizer (required)

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
≥ 10 GB	q8_0
6–9 GB	q6_k
CPU only	f16 (slow)

---

Quick start

# 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

./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.

---

All CLI options

Flag	Default	Description
-m, --model	model.gguf	Path to GGUF model file
-t, --tokenizer	tokenizer.json	Path to tokenizer.json
-text	"Hello world"	Text to synthesize
-pa, --prompt-audio	—	Reference audio (WAV/MP3)
-pt, --prompt-text	—	Transcript of reference audio
-o, --output	out.wav	Output WAV path
-v, --vulkan	-1 (CPU)	Vulkan device index
-threads N	4	CPU threads
-max-tokens N	512	Max tokens (~21s per 440 tokens)
-temp F	0.7	Sampling temperature
-top-p F	0.7	Top-p nucleus sampling
-top-k N	30	Top-k sampling

---

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

Attribution: "This model is licensed under the Fish Audio Research License, Copyright © 39 AI, INC. All Rights Reserved."

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