MAGNeT Medium 10s

model·generative music model·active

Non-autoregressive MAGNeT checkpoint for 10 seconds of music.

Verified facts

Official site	https://huggingface.co/facebook/magnet-medium-10secs ↗
Geography	unknown
parameter count	1.5B
huggingface repo	facebook/magnet-medium-10secs
api	unknown
inputs	text
license	CC-BY-NC-4.0 model weights; MIT code
outputs	music
weights	public
modalities	music; text
dependencies	Meta EnCodec tokenizer
model family	MAGNeT
access method	downloadable weights; AudioCraft API
exact version	magnet-medium-10secs
repository links	https://huggingface.co/facebook/magnet-medium-10secs; https://github.com/facebookresearch/audiocraft
evaluation results	result: FAD: 4.61 · KLD: 1.14 · text consistency: 0.28 · source: https://huggingface.co/facebook/magnet-small-10secs · benchmark: MusicCaps
evidence separation	developer claims: · measured results: Model card reports {'FAD': 4.61, 'KLD': 1.14, 'text_consistency': 0.28} on MusicCaps for this checkpoint. · independent evaluations:
developer maintainer	Meta AI
material us relevance	Publicly accessible in or materially offered to the United States market as of snapshot.
documented capabilities	text-to-music; fixed generation length: 10 seconds
products exposing model	unknown
commercial use conditions	Model weights are non-commercial under CC-BY-NC-4.0.
known geographic or policy restrictions	Non-commercial weights
safety watermarking disclosure provenance	unknown

introduces (by)	Masked Audio Generation using a Single Non-Autoregressive Transformer 2024-01-09 — now
depends on	Meta EnCodec 32 kHz (MusicGen checkpoint)