Skip to content
← Models
Open weightsMeta (FAIR)/10 trustVerified
MusicGen-Large logo

MusicGen-Large

Meta's open text-to-music model — describe a song, get 12 seconds of music.

Visit official page →

About

MusicGen-Large (3.3B parameters) is Meta FAIR's text-to-music model: a single-stage autoregressive transformer language model operating over discrete audio tokens from a 32kHz EnCodec tokenizer, conditioned on text descriptions or melodic features (an optional reference audio guides the harmony). It was trained April-May 2023 on 20,000 hours of licensed music (10,000 high-quality internal tracks plus Shutterstock and Pond5), generates mono and stereo clips up to ~12 seconds, and is distributed with code under MIT and weights under CC-BY-NC 4.0 (non-commercial).

There is no independent public benchmark leaderboard for music generation. Meta's provider-reported evaluations found MusicGen superior to the evaluated baselines (MusicLM, Riffusion, Mubert, Noise2Music) on objective metrics (FAD, KL divergence) and human-listening scores (overall quality OVL, text relevance REL) on the MusicCaps benchmark.

Specs

Context window

N/A — text prompts plus an optional ~30s melody-reference audio; generates clips up to ~12s

Pricing

Free (open weights; non-commercial license)

Free quota

Free — self-host; Hugging Face Spaces demos available. Weights are CC-BY-NC 4.0: research and non-commercial use only.

API access

Self-host; Hugging Face Spaces demos and inference endpoints

Weights

Open weights

Released

June 8, 2023

Benchmarks

No independent public leaderboard exists. Provider-reported (Meta, MusicCaps evals):

  • Outperformed MusicLM, Riffusion, Mubert, Noise2Music on objective metrics (FAD, KL)
  • Outperformed the same baselines on human-listening scores (overall quality OVL, text relevance REL)

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

text-to-music generationmelody-conditioned music creationgame and film scoring prototypesmusic-generation research

Trust score — 6/10

/10 trustVerified

Scored 6/10 (verified tier, metrics-v1). Evidence: HF: facebook/musicgen-large; HF downloads 84,442 (peer-normalized); released 1223 days ago. Manual review: no independent benchmark leaderboard exists for music gen; all head-to-head claims are Meta's own evals. License: CC-BY-NC 4.0 — commercial-use ban (hard restriction).

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
7.6/10
Maintenance · recency, releases, issue load
4.0/10
Openness · license + docs
5.8/10
Manual review · sanity read + red-flag check
7.0/10
  • · HF: facebook/musicgen-large
  • · HF downloads 84,442 (peer-normalized)
  • · released 1223 days ago
  • · manual review: no independent benchmark leaderboard exists for music gen; all head-to-head claims are Meta's own evals

More in the musicgen family

Sources