This is not a full-stack pure MLX audio pipeline yet: token generation runs with MLX, while FLAC decoding currently uses the official PyTorch Flow1dVAE / separate-tokenizer bridge in ailuntx/SongGeneration-MLX.

TL;DR


Variant	`v2-medium`
Precision	`fp32`
Converted component	SongGeneration `audiolm` token generator
Runtime	`ailuntx/SongGeneration-MLX`
Official model	`tencent/SongGeneration`
Official code	`tencent-ailab/songgeneration`

Quick Start

bash

git clone https://github.com/ailuntx/SongGeneration-MLX.git
cd SongGeneration-MLX
python -m venv .venv
.venv/bin/pip install -e .
.venv/bin/pip install -U huggingface_hub hf_transfer

HF_HUB_ENABLE_HF_TRANSFER=1 .venv/bin/hf download mlx-community/SongGeneration-v2-medium-fp32 --local-dir ./models/SongGeneration-v2-medium-fp32

.venv/bin/python -m songgeneration_mlx.cli \
  --model ./models/SongGeneration-v2-medium-fp32 \
  --lyrics '[verse] hello from mlx [chorus] sing it again' \
  --description 'Pop, female vocal, bright production, [Musicality-medium].' \
  --duration 2 \
  --top-k 50 \
  --temperature 0.9 \
  --output tokens_2s.npz

Decode Tokens

The MLX runtime writes discrete song tokens. To create FLAC audio, use the official decoder bridge in ailuntx/SongGeneration-MLX. The bridge needs the official SongGeneration runtime/ assets, but it does not need the original SongGeneration LM model.pt when --mlx-model is used.

bash

python -m venv .venv-decoder
.venv-decoder/bin/pip install -U pip
.venv-decoder/bin/pip install \
  -r third_party/SongGeneration/requirements.txt \
  -r third_party/SongGeneration/requirements_nodeps.txt \
  soundfile

HF_HUB_ENABLE_HF_TRANSFER=1 .venv/bin/hf download tencent/SongGeneration \
  --include "runtime/*" \
  --local-dir ./third_party/SongGeneration

PYTORCH_ENABLE_MPS_FALLBACK=1 SONGGEN_DEVICE=mps \
.venv-decoder/bin/python scripts/decode_tokens_official.py \
  --mlx-model ./models/SongGeneration-v2-medium-fp32 \
  --tokens ./tokens_2s.npz \
  --output ./output_2s.flac \
  --device mps

Variants

Variant	Disk	Notes
`SongGeneration-v2-medium-fp32`	10G	high-precision medium baseline
`SongGeneration-v2-medium-bf16`	5.2G	recommended medium bf16 quality baseline
`SongGeneration-v2-medium-8bit`	2.8G	smaller medium checkpoint
`SongGeneration-v2-medium-4bit`	1.5G	smallest medium checkpoint
`SongGeneration-v2-large-fp32`	19G	high-precision large baseline
`SongGeneration-v2-large-bf16`	9.5G	large bf16 quality baseline
`SongGeneration-v2-large-8bit`	5.0G	smaller large checkpoint
`SongGeneration-v2-large-4bit`	2.7G	smallest large checkpoint
`SongGeneration-v2-fast-*`	pending	upstream fast weights were not publicly available when checked on 2026-05-31

Layout

text

SongGeneration-v2-medium-fp32/
|-- model-00001-of-000xx.safetensors
|-- model.safetensors.index.json
|-- config.json
|-- mlx_manifest.json
|-- config.official.yaml
|-- vocab.yaml
`-- qwen2_tokenizer/

Validation

Local Apple Silicon validation was run on the medium bf16 path:

Test	Result
12s MLX token generation	550 pattern steps, about 1 minute wall time
12s official decoder bridge	73.27s wall time
12s FLAC	48kHz stereo, 12.000s, RMS about `0.163`

The official recent-token repetition penalty is implemented in the MLX runtime. Without it, 12s generations collapse into repeated tokens and decode close to silence.

License

License follows the upstream SongGeneration release. Check the official model card and repository for the authoritative model license.

Citation

bibtex

@misc{songgeneration-mlx,
  title  = {SongGeneration-MLX: Apple MLX port of SongGeneration},
  author = {ailuntx},
  year   = {2026},
  url    = {https://github.com/ailuntx/SongGeneration-MLX},
}

@article{lei2025levo,
  title   = {LeVo: High-Quality Song Generation with Multi-Preference Alignment},
  author  = {Lei, Shun and Xu, Yaoxun and Lin, Zhiwei and Zhang, Huaicheng and Tan, Wei and Chen, Hangting and Yu, Jianwei and Zhang, Yixuan and Yang, Chenyu and Zhu, Haina and Wang, Shuai and Wu, Zhiyong and Yu, Dong},
  journal = {arXiv preprint arXiv:2506.07520},
  year    = {2025},
}