MLX port of Meta's MAGNeT — a masked parallel text-to-music model — quantized to INT4 weight-only for on-device generation on Apple Silicon. EnCodec 32 kHz audio decoder; T5-base text encoder; per-codebook iterative decoding with restricted-context self-attention.

Model


Parameters (LM)	1.5B
Quantization	INT4 weight-only, group size 64
Format	MLX safetensors
Sample rate	32 kHz mono
Output length	30 s per generation (fixed)
Decoding steps	50 total (`[20, 10, 10, 10]` across 4 RVQ codebooks)
Bundle size	1357 MB on disk
Source	facebook/magnet-medium-30secs

Performance (Apple Silicon, 30 s audio)

Metric	Value
Real-time factor (wall / audio)	1.20
Peak RSS	2215 MB
CLAP score (laion/clap-htsat-unfused, 5 prompts)	0.334

Usage

python

from huggingface_hub import snapshot_download
bundle = snapshot_download("aufklarer/MAGNeT-Medium-30secs-MLX-4bit")
# See https://github.com/soniqo/speech-swift for production usage.

Source

License

CC-BY-NC 4.0 — inherited from upstream MAGNeT weights. Non-commercial use only.