MiniMax Music Generator

MiniMax Music 3 is an AI music model from MiniMax, the lab behind the MiniMax-H3 video model. MiniMax released it with open weights on Hugging Face as MiniMaxAI/MiniMax-Music3. It writes complete songs up to five minutes long from two things you give it: the lyrics to sing, and a description of the music.

What sets it apart is that it sings your words. Most music generators write their own lyrics from a prompt; MiniMax Music 3 performs the lyric sheet you wrote, section by section, in the order you tag it: an intro, verses, a pre-chorus, choruses, a bridge, a solo and an outro. The description sets everything else, from the genre, tempo and key to the singer and how the arrangement grows from the first bar to the last. It also makes instrumentals, from film scores to study beats.

Mage runs MiniMax Music 3 in your browser, next to the rest of its image, video and audio models, so there is nothing to download and no graphics card to own. Mage is an independent creative platform, not part of MiniMax. What it adds is everything around the song: unlimited plans, LoRAs on top of the model, a private library of everything you make, and video models that can turn the track into a music video.

MiniMax Music 3 reads two inputs, and each has one job. The prompt describes the music: the genre and subgenre, the tempo in beats per minute, the key, the mood and how it changes, who sings and how, and what the instruments do. The lyrics are the words to sing, and nothing else. The description is never sung, and the lyrics never set the style, so a song comes out best when each stays in its lane.

In the lyrics, section tags give the song its shape. Write [intro], [verse], [pre-chorus], [chorus], [bridge], [solo] or [outro] alone on a line, with the sung lines below it. A tag with nothing under it, such as a [solo] before the last chorus, leaves room for an instrumental passage. One rule matters more than any other: words written on the same line as a tag are dropped, so always start a new line after the tag.

The MiniMax Music 3 prompting guide has a fill-in template for the description, tagged lyric sheets, and worked examples in several genres.

• Section tags alone on their lines
• Original lyrics, one sung line per line
• Repeat the chorus in full each time it returns
• An [outro] so the song ends instead of stopping

MiniMax recommends a description in three parts. First, the global picture: the genre and subgenre, the tempo, the key and scale, the emotional arc from the first bar to the last, a scene the song would play in, and how the mix should sound. Second, the vocals: the singer's gender, timbre and register, how they deliver the verses and the chorus, any harmonies or backing vocals, and effects such as reverb or delay. Third, the arrangement, as a timeline: what plays in the intro, what enters at the first chorus, what drops out in the bridge, and how the song ends.

Always describe the singer. MiniMax notes that a song whose description leaves out the voice can drift into an instrumental. And decide the details yourself: whatever you leave open, the model fills with its most average choice, so "acoustic pop" gives a generic arrangement, while "indie folk pop ballad, 84 BPM, G major, fingerpicked guitar and felt piano in the verses, cello and brushed drums at the first chorus" gives the song you hear in your head.

For a song whose lyrics the model writes from a description, use Eleven Music instead. MiniMax Music 3 is the model for lyrics you wrote yourself.

• Genre, tempo and key first
• Name the singer and the delivery
• Lay out the arrangement section by section
• Describe the mix: warm, wide, raw or polished

The Length setting is the longest the song may run, up to five minutes. It is a limit, not a target: MiniMax Music 3 ends the song on its own when the lyrics and the arrangement are done. A short lyric sheet gives a short song, whatever the Length, and a long lyric sheet with a short Length is cut off at the limit, even in the middle of a line.

So plan the lyrics for the length you want. About a verse and a chorus fill a minute; two verses, three choruses and a bridge fill three or four. A practical workflow is to try the description with a short Length and one verse and chorus, keep the seed when the sound is right, and then render the full song.

For music with no vocals, leave the lyrics empty or turn on Instrumental, and spend the whole description on the music. Name the instruments in the order they enter, the key and any key change, the room, and how the piece ends. A cinematic score that opens on a lone cello and builds to a full orchestra, or a lo-fi study beat with dusty electric piano and vinyl crackle, both come out of the same model that sings.

For short sound effects and ambience, Seed Audio is the better tool, and the AI music generator page covers beds and loops: see the AI music generator.

Because MiniMax Music 3 has open weights, its community trains LoRAs for it: small add-on files that push the model toward a genre, a production style or a vocal character. Some change how the song is planned, others how it sounds. On Mage they work on top of the model. Paste the Hugging Face or CivitAI link of a MiniMax Music 3 LoRA on the Import page, and it appears in the LoRA picker for MiniMax Music 3.

Start a new LoRA at weight 1, and lower it if the song starts to lose the lyrics or the structure. When a LoRA was trained with a trigger word, put that word in the description.

Open weights mean anyone can download MiniMax Music 3 from Hugging Face and run it themselves. In practice, an 8-billion-parameter language model paired with a flow-matching audio model takes a powerful graphics card, tens of gigabytes of model files and a working setup, and a five-minute song keeps that card busy for minutes.

On Mage it runs in the browser on any device, a phone included, with LoRA support already in place. And because Mage carries video models in the same account, a song you make here can become the soundtrack of a clip: the AI music video generator turns a track into a performance by your saved characters.

MiniMax Music 3 is priced by the Length you pick, the longest the song may run. On a paid plan, a song runs without Gems when its length fits what the plan covers, and higher plans cover longer songs. Unlimited runs use the standard queue.

Without a plan, or for a song longer than your plan covers, MiniMax Music 3 runs on Gems, and the price shows before you generate. Gem packs start at $10 for 10,000 Gems, and every paid plan includes monthly Gems, 1,500 on Pro.

Compare what each plan covers on the membership page, or see the MiniMax Music 3 model page for the plan that covers each song length.