Fishaudio S2 Pro MLX 4bit

--- library_name: mlx-audio tags: - mlx - mlx-audio - text-to-speech - speech - speech-generation - voice-cloning - tts - fish-audio - s2-pro - quantized - 4bit - fish_qwen3_omni base_model: fishaudio/s2-pro license: other license_name: fish-audio-research license_link: https://huggingface.co/fishaudio/s2-pro/blob/main/LICENSE pipeline_tag: text-to-speech language: - en - zh - ja - ko - es - pt - ar - ru - fr - de --- # Fish Audio S2 Pro MLX 4-bit MLX 4-bit quantization of [fishaudio/s2-pro](https://huggingface.co/fishaudio/s2-pro), produced with [mlx-audio](https://github.com/Blaizzy/mlx-audio) on Apple Silicon. **Built with Fish Audio.** ## Provenance Converted directly from `fishaudio/s2-pro` bf16 safetensors via `python -m mlx_audio.convert ... -q --q-bits 4 --q-group-size 64`. The `codec.pth` audio decoder, tokenizer, and chat template are preserved alongside the quantized LLM weights. ## Quantization - Method: MLX affine quantization, group_size=64, 4-bit - Actual bits per weight after conversion: **~5.0** (some small projections are skipped and remain bf16 — this is expected behavior for mlx.nn.quantize on multimodal models) - Size on disk: ~2.4 GB (vs ~8 GB for bf16) - Target: Apple Silicon with ≥ 16 GB unified memory (comfortable on M-series with 24 GB+) ## Architecture `fish_qwen3_omni` — a Qwen3-based decoder-only LLM that emits audio tokens consumed by the `codec.pth` DualAR codec. The text side is quantized; the codec is kept as-is. ## Quickstart ```python from mlx_audio.tts import load model = load("majentik/fishaudio-s2-pro-MLX-4bit") audio = model.generate( text="Hello, this is a test of Fish S2 Pro at 4-bit MLX.", voice="en_default", ) ``` Fish S2 Pro supports zero-shot voice cloning from a reference audio sample; see the base model card for the reference-conditioning API. ## License & Use Restrictions This model is a derivative work of **Fish Audio S2 Pro**, licensed under the **Fish Audio Research License**. - **Research + Non-Commercial use:** free, subject to the license. - **Commercial use:** requires a separate written agreement with Fish Audio — see https://fish.audio or email `[email protected]`. - **Use restriction:** outputs must not be used to create or improve any foundational generative AI model (excluding the Model itself or derivatives of it). - **Attribution:** downstream use must include this NOTICE an...

Source context: 58 downloads · 0 likes · Pipeline text-to-speech · Library mlx-audio · Repo 3z1mmaj3/fishaudio-s2-pro-MLX-4bit

Source context: 77 downloads · 0 likes · Pipeline text-to-speech · Library mlx-audio · Repo 3z1mmaj3/fishaudio-s2-pro-MLX-4bit