mlx-community/MOSS-TTS-Local-Transformer-v1.5-bf16

MLX conversion of OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5 (OpenMOSS / MOSI.AI) — the 4B MossTTSLocal speech model (Qwen3 backbone + a local transformer emitting 12 RVQ codebooks per frame), zero-shot voice cloning, streaming synthesis, 48 kHz stereo output through MOSS-Audio-Tokenizer-v2.

Precision: bf16 (unquantized, weights as upstream) — 8.5 GB. Converted with mlx_audio.convert (model type moss_tts_local) from the upstream safetensors, tokenizer files copied verbatim; the upstream remote-code .py files are not needed by MLX and were dropped.

Sibling tiers: MOSS-TTS-Local-Transformer-v1.5-bf16 · -8bit · -4bit. The codec is mlx-community/MOSS-Audio-Tokenizer-v2-bf16 (this config's audio_tokenizer_name_or_path still points at the upstream fp32 repo, which mlx-audio also loads directly).

Use with mlx-audio

from mlx_audio.tts import load
from mlx_audio.audio_io import write as audio_write

model = load("mlx-community/MOSS-TTS-Local-Transformer-v1.5-bf16", lazy=True)
result = next(model.generate(
    text="Hello, this is MOSS-TTS running on MLX.",
    ref_audio="speaker.wav",          # optional zero-shot cloning prompt
    language="English",
    max_tokens=400,
))
audio_write("out.wav", result.audio, result.sample_rate)   # 48 kHz stereo

Streaming (stream=True, streaming_interval=2.0) is supported on this checkpoint; see the mlx-audio moss_tts README. Fixed 12-codebook RVQ depth — do not pass a different n_vq_for_inference.

License

Apache-2.0 (upstream weights and this conversion). Please credit OpenMOSS for the model.