minimax/music-3

MiniMax Music 3 writes and performs a complete song from lyrics and a style description.

$0.002 / second
GPU: A100

Model Inputs

string

Music description: style, mood, vocals, instrumentation and arrangement. For precise control use a Structured Caption with global metadata (genre, BPM, key, emotional progression), vocal details, and a section-by-section arrangement.

string

The lyrics to sing. Structure tags such as [intro], [verse], [pre-chorus], [chorus], [post-chorus], [bridge], [instrumental], [solo] and [outro] must each be on their own line; text on the same line as a leading tag is dropped by the model's input contract.

number

Upper bound on the generated audio length in seconds. The model may stop earlier; the actual duration is returned in the output.

integer

Random seed for reproducibility. If not provided, a random seed will be used.

integer

Number of flow-matching Euler steps per 8-second denoising chunk. More steps improve quality at the cost of speed.

number

Classifier-free guidance scale of the flow-matching stage.

Result

Idle
Waiting for your input...

Pricing

minimax/

music-3

You are billed for what this model actually produces — by the second — so a shorter request costs less and you never pay for idle time.

Rate
$0.002 / second
or
500 seconds / $1

For example, a request of 60 seconds costs about $0.12.

You are charged for what the request actually produces, so the final amount scales with the seconds you use.

ProviderPrice ($)Saving (%)
Synexa$0.12-
fal$0.12same price

Comparison prices are for an equivalent request.

Readme

MiniMax Music 3 generates full songs of up to five minutes from two inputs: the lyrics to sing, and a description of the music.

Structure tags in the lyrics - [intro], [verse], [pre-chorus], [chorus], [bridge], [instrumental], [solo], [outro] - must each sit on their own line. The prompt should cover genre, mood, vocals, instrumentation and arrangement; for tight control write a structured caption with BPM, key and a section-by-section breakdown.

duration is an upper bound - the model may stop earlier, and the real length comes back in the output.