Drama 3 MLX bf16

Built with Fish Audio.

Fish Audio Drama 3 converted for mlx-audio on Apple Silicon, with nothing quantized. This is an unofficial community conversion; Fish Audio did not release or endorse it.

Converted from the updated weights Fish Audio released on 2026-10-29 (revision 78e5ab7). The conversion of the first release stays at revision d947126 of this repo.

  • routed experts: dequantized from the block-FP8 release to bfloat16
  • every other weight: bfloat16, as released
  • 67 GB on disk, about 74 GB peak memory while generating, so it needs a Mac with at least 96 GB

For smaller Macs, mlx-community/fish-audio-drama-3-4bit needs about 26 GB.

Usage

Use mlx-audio 0.6.2 or later, which samples Drama 3 with SGLang-Omni's defaults:

pip install -U "mlx-audio>=0.6.2"
from mlx_audio.tts.utils import load_model

model = load_model("mlx-community/fish-audio-drama-3-bf16")

for result in model.generate(
    text="[sighs] Fine. If we miss it, it's on you.",
    ref_audio="reference.wav",
    ref_text="Transcript of the reference clip.",
    stream=True,
    first_chunk_frames=4,
):
    print(result.audio.shape)  # 44.1 kHz chunks
python -m mlx_audio.tts.generate --model mlx-community/fish-audio-drama-3-bf16 \
    --text "[sighs] Fine. If we miss it, it's on you." \
    --ref_audio reference.wav --ref_text "Transcript of the reference clip."

Speaker tags (<|speaker:0|> to <|speaker:15|>), multiple references and streaming are described in the model docs.

Speed

On an M3 Max with 128 GB:

  • about 1.4x real time
  • first streamed audio in 270 to 300 ms once a voice is cached
  • loading takes about 12 s from a warm page cache, and the first call after loading about 3 s

Conversion

Made from fishaudio/drama-3 at revision 78e5ab7 with mlx-audio's converter, with nothing quantized:

python -m mlx_audio.tts.models.fish_drama3.convert --src drama-3 --dst drama-3-mlx-bf16 \
    --expert-bits 0 --dense-bits 0 --tower-bits 0

NOTICE lists every change to the original weights.

License

The Fish Audio Community License applies, including its use restrictions and its terms for commercial use and hosted services.

Downloads last month
614
MLX
Hardware compatibility
Log In to add your hardware

Quantized

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

Model tree for mlx-community/fish-audio-drama-3-bf16

Finetuned
(8)
this model