MusicGen Large
Audio Available Full comparison ↗ meta/musicgen-large · by Meta AI
· transformer-decoder
Pricing — 1 offering(s)
About this offering: Meta's MusicGen Large, a 3.3-billion-parameter open-weights music model (weights under CC BY-NC 4.0, so not for commercial use). It generates instrumental music, with no vocals, from a text prompt or, in the Melody variant, a reference melody. On Replicate it runs on an A100 GPU and is billed by compute time, so the cost per run varies with the settings; the price shown is Replicate's own per-run estimate. You can also self-host it.
Per generation (Replicate-hosted)
- $0.042 / generation Historical 2023-08-01 → 2026-09-19
- $0.10 / generation
Current
2026-09-20 → present
Replicate's own rolling estimate: approximately $0.10 per run (about 10 runs per $1) on an Nvidia A100 80GB…
Replicate's own rolling estimate: approximately $0.10 per run (about 10 runs per $1) on an Nvidia A100 80GB, roughly 72 seconds per prediction, for both the Large and Melody variants. It varies with inputs (Replicate showed $0.036 on 2026-09-30), so treat it as an estimate rather than a list price.
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how MusicGen Large fits into a cost-aware routing setup
See how →Capability profile
How MusicGen Large rates across core capability dimensions, with the task-level evidence behind each rating.
Good audio quality for an open-weights model; not at the level of Suno v4 or Stable Audio 2.0 on instrumental fidelity. No post-processing artifacts from a well-configured setup.
Strong genre coverage from training data. Text conditioning works well for mainstream and classical genres; niche sub-genres are less reliable.
No vocal generation. Pure instrumental output only — a fundamental architectural limitation, not a configuration option.
Default generation limit of 30 seconds per call. Longer audio requires multiple generations and stitching.
No stem export. Mixed-down audio output only.
Benchmarks
| Benchmark | Score | Config | Source |
|---|---|---|---|
| FAD (Fréchet Audio Distance, published in paper) | — | Copet et al. report FAD and KL-divergence scores on MusicCaps dataset. MusicGen Large outperforms prior open models (MusicLM, Riffusion). Exact scores in the paper. | source ↗ |
Operator guidance
Choose MusicGen Large when open-weights ownership matters (CC BY-NC 4.0 allows commercial use with attribution), or when self-hosting eliminates per-call costs at scale. For production quality with vocals, use Suno v4. For longer-form high-fidelity instrumentals, use Stable Audio 2.0. For melody conditioning specifically, MusicGen's melody-conditioning feature is unique among major models.
Use cases
- Background music generation in applications where open-weights licensing matters
- Research and experimentation on music generation pipelines
- Self-hosted music generation to eliminate per-call API costs at scale
- Melody-conditioned generation: constraining output to follow a reference melody
Limitations
- Instrumental only — no vocal generation
- 30-second generation ceiling per call; longer clips require stitching
- CC BY-NC 4.0 licence — commercial use requires attribution; no-commercial clause applies to the model weights
- Requires GPU infrastructure for reasonable generation speed (no first-party hosted API)