MAGNeT
Text-to-music in a fraction of a second — up to 7x faster than autoregressive baselines
About
MAGNeT (Masked Audio Generation using a Single Non-Autoregressive Transformer) is Meta FAIR's answer to slow audio generation. Instead of decoding token by token, one transformer predicts spans of masked audio tokens in parallel over a 32 kHz EnCodec tokenizer with 4 codebooks at 50 Hz, with an external rescoring step to clean up fidelity. It ships in 300M and 1.5B sizes with separate text-to-music and text-to-sound variants, trained on 16,000 hours of licensed music (an internal 10K-track set plus Shutterstock and Pond5 data). Code is MIT and weights are CC-BY-NC 4.0, so research and non-commercial use is fine but commercial products need another path. Generations are fixed 10- or 30-second clips, and the medium model wants a 16 GB GPU.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported: quality parity with baselines on Frechet Audio Distance at roughly 7x the generation speed (paper)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Meta model card verified: trained Nov 2023–Jan 2024, weights CC-BY-NC 4.0, code MIT; 7x speed claim is team-reported from the paper — labeled as such; Exact public release day not verified — releasedAt null.