
Melody conditioning — hum an idea and get an arrangement
Supplying a melody specifies the actual tune rather than a style, which is a fundamentally more musical form of control than any prompt.
Meta AI
Hum an idea and get it arranged — the original melody-conditioned model.
In short
Figures verified 2026-08-10. This field moves quickly — re-check before relying on them.
MusicGen is the model most people in this field learned on, and melody conditioning is still the feature that makes it worth using rather than merely studying. Text prompts describe a style; a hummed melody specifies the actual tune, which is a fundamentally different and more musical kind of control. If you can hum the idea in your head, you can have it arranged, and no amount of adjectives gets you there otherwise. The variant family matters practically: smaller versions run on modest hardware for prototyping, larger ones produce noticeably better audio on a real GPU, and the same code path serves both. Documentation and community support are the best in open music generation by a wide margin, which means problems you hit have usually been solved publicly already. It has been overtaken on raw audio quality by newer models and remains the most instructive place to start.
Strengths

Supplying a melody specifies the actual tune rather than a style, which is a fundamentally more musical form of control than any prompt.

Model sizes from small to large share a code path, so you can prototype on modest hardware and scale up without rewriting anything.

Documentation and community answers are far ahead of the rest of open music generation, so most problems you hit are already solved publicly.
How it compares
MusicGen is Meta AI's open-weight music generation model, distinguished by melody conditioning: you supply a melody, hummed or played, and it generates an arrangement that follows it. It ships in several sizes and is the best-documented open model in the field.
Compare with MusicGenerate
Searched as musicgen, meta musicgen, audiocraft, musicgen melody and facebook ai music generator. Melody conditioning is what people are looking for when they arrive from the last of those.
Supplying a melody — hummed or played — and getting an arrangement that follows it. Text prompts describe a style; a melody specifies the actual tune, which is a fundamentally more musical kind of control.
Smaller variants run on modest hardware for prototyping and larger ones produce noticeably better audio on a real GPU. They share a code path, so you can scale up without rewriting anything.
Read the specific licence first. The code and weights carry Meta's research-oriented licensing, which restricts some commercial use — the most common misunderstanding about MusicGen is assuming open weights means commercially unrestricted.
It has been overtaken on raw audio quality and remains the most instructive place to start, with the best documentation and community support in open music generation by a wide margin.
The weights are free to download; the licence is the constraint rather than the price. Meta's research-oriented terms restrict some commercial use, which is the most common misunderstanding about it.
Smaller variants run on modest hardware for prototyping; the larger ones want a real GPU. They share a code path, so you can develop on one and deploy on another.
Keep exploring
More models
Models and platforms we track, compared on capability and licence.
Hum an idea and get it arranged — the original melody-conditioned model.