MiniMax Music 3.0: Songs That Keep Their Shape
MiniMax Music 3.0 is built for the part after the hook: full songs whose second chorus still sounds like the first. Here's how it works and how to brief it on each::labs.
By Eftal Yurtseven · Oct 28, 2026 · 7 min read
Thirty seconds of AI music is a solved problem. A hook, a loop, a vibe. The trouble starts at minute two, when the second chorus arrives and the model has quietly forgotten what the first one sounded like. The singer changes. The drums wander off. What should be a song turns into a playlist of fragments that happen to share a file.
MiniMax Music 3.0 is built for the part after the hook. MiniMax released it on July 16, and its model card describes complete songs up to five minutes long that keep their themes, rhythm, vocal identity and arrangement intact from intro to outro. It's live on each::labs now, and it's the most interesting music model in the MiniMax Music family for anyone who needs a whole track, not a sketch.
What MiniMax Music 3.0 Actually Is
Music 3.0 is a lyrics-plus-description model. You give it the words to sing and a written brief for how the song should sound, and it arranges, performs and mixes a finished track. MiniMax's own summary of the upgrade names three things: better understanding of creative intent, so less of the usual AI drift; a big jump in sound quality, with specific instruments and real playing techniques like slides and legato; and a new vocal engine with control over melody, pronunciation, breathing and layered harmonies.
The architecture explains the long-form claim. According to the MiniMax Music 3 model card, an 8B "global" language model plans the song's long-range structure frame by frame, while a smaller 0.6B "local" model fills in fine acoustic detail inside each frame. One part of the system thinks about where the song is going. The other thinks about how this exact moment sounds. That split is why a bridge can still feel like it belongs to the verse that came three minutes earlier.
MiniMax also published the open weights on Hugging Face, which is rare for a model at this level. On each::labs you don't need to touch any of that: it's one request.
Lyrics Are the Score, So Write Them Like One
Most people paste lyrics in as a block of text. That's leaving control on the table. The lyrics field on each::labs accepts up to 3,500 characters with line breaks and section tags: [Intro], [Verse], [Pre Chorus], [Chorus], [Bridge] and [Outro]. MiniMax's model card lists more, including [Instrumental], [Solo] and [Post-Chorus].
Those tags are the song's floor plan. A chorus tag tells the model this section comes back, should hit harder and should sound like itself every time. A bridge tag tells it to change something. If you want a guitar solo between the second chorus and the bridge, write a [Solo] section instead of hoping. And keep choruses word-for-word identical when they repeat. If you change one line, you're telling the model it's a different section.
[Intro]
[Verse]
Headlights on the harbor road
Salt still in my coat
[Pre Chorus]
Every mile a little slower
[Chorus]
Bring me home before the tide turns
Bring me home before the light burns out
[Solo]
[Bridge]
I kept the map you drew in pencil
[Chorus]
Bring me home before the tide turns
Bring me home before the light burns out
[Outro]
Write the Style Prompt Like a Production Brief
The prompt field (up to 2,000 characters on each::labs) carries style, mood, instruments, voice and scenario. "Sad pop song" will get you a sad pop song. It won't get you your sad pop song.
MiniMax's model card recommends a structured caption with three parts, and it's the best advice in the whole release. Global metadata first: genre, subgenre, BPM, key, scale, emotional progression, listening scenario and production profile. Then vocal details: gender, timbre, performance style, harmonies, backing vocals and vocal effects. Then arrangement: lead and supporting instruments, how they change section by section, groove, bass, percussion, textures and space.
Indie folk rock, 92 BPM, D major, starts wistful and builds to hopeful by the last chorus. Late night drive. Warm, slightly roomy analog production. Vocals: male baritone, close and breathy in the verses, full voice in the choruses, soft harmony on the final chorus. Arrangement: fingerpicked acoustic guitar and brushed drums in verse one, bass and organ enter at the pre chorus, electric guitar slide solo before the bridge, everything drops to acoustic guitar for the bridge, full band for the last chorus.
Notice the arrangement line describes change over time. That's the part most prompts skip, and it's the part Music 3.0 is built to follow. A brief that only names instruments gets you a static mix. A brief that says when they enter gets you a song that builds.
Instrumentals and Songs Without Lyrics
Not every job needs a singer. Set is_instrumental to true and Music 3.0 drops the vocals, which makes it useful for background beds under ads, explainer videos, game menus and podcasts. Keep the section tags in your head even here: describe the shape in the prompt ("quiet intro, groove enters at 0:20, breakdown in the middle, big final section") so the bed has places for an edit to land.
If you have a mood but no words, turn on lyrics_optimizer. When the lyrics field is left empty, MiniMax can write lyrics from your style prompt. It's a fast way to get a demo for a pitch. For anything that'll carry your brand, write or at least edit the words yourself: lyrics are the part listeners remember, and they're the part you'll want to own.
Where MiniMax Music 3.0 Earns Its Place
Video teams use it for original soundtracks that run as long as the cut. A two-minute brand film can have a track that actually has a structure (intro under the logo, a lift where the product appears, an outro under the end card) instead of a loop faded out at the end. Game studios can draft menu themes and level music in several moods before a composer refines the winners.
Creators and social teams get the most out of the vocal side. A jingle with real lyrics, a parody, a song about the product written in the brand's voice: those are hard to source quickly and easy to brief. Agencies can spin variations of one track for different markets by changing the language of the lyrics and the genre line of the prompt while keeping the structure tags identical.
And for app builders, it's a full-song engine behind a single endpoint. A user writes a few lines and picks a mood. Your app sends lyrics and a prompt and gets a finished track back.
Getting Better Songs Out of MiniMax Music 3.0
Start short. Write a verse and a chorus, generate, and listen for whether the voice and groove are right before you commit a full five-section song to the run. Once the sound is right, add the rest of the structure and keep the prompt unchanged.
Describe the voice like a casting note, not an adjective. "Female alto, slightly husky, conversational in verses, belted chorus, doubled vocal on the last line" beats "great female vocal." MiniMax specifically calls out control over breathing and harmonies, so ask for them.
Give the song an emotional arc. The model card lists emotional progression as a metadata field for a reason. "Starts restrained, opens up at the second chorus, ends quiet" gives the arrangement a direction to follow.
Pick your audio settings for the destination. The audio_settings object defaults to MP3 at 44.1 kHz and 256 kbps, and also offers WAV, lower sample rates and lower bitrates. Use WAV when the track is going into an editor, MP3 when it's going straight to a player.
MiniMax Music 3.0 vs Other Music Models on each::labs
MiniMax's own line is still there if you need it: MiniMax Music 2.5 and Music v2 remain on each::labs, but 3.0 is the one to start with for anything long or vocal-led. Google's Lyria 3.5 also does full songs, and our Lyria 3.5 write-up shows where it differs. Mureka V9.5 is the other strong song model, covered in Mureka V9.5: songs that hold together.
The honest way to choose is to run the same lyrics and the same brief through two of them and listen. On each::labs that's a model id change, not a new integration, and you can chain the winner with a video model in an each::labs flow so the soundtrack arrives with the cut. Browse the full model catalog for voice and sound models to pair it with.
Frequently Asked Questions
How long can MiniMax Music 3.0 songs be?
MiniMax's model card describes full songs up to five minutes, with intro, verses, choruses, bridge, instrumental breaks and outro in one piece. Length follows your structure, so the number of sections you write in the lyrics shapes how long the track runs.
Can MiniMax Music 3.0 make instrumental music?
Yes. Set is_instrumental to true and it generates the track without vocals. Describe the arrangement and its changes over time in the prompt so the instrumental still has a shape.
Do I need to write lyrics for MiniMax Music 3?
No. Turn on lyrics_optimizer and leave the lyrics empty, and MiniMax writes them from your style prompt. For finished work, editing the lyrics yourself gives you far more control over the result.
What should go in the MiniMax Music 3.0 prompt?
Think production brief: genre, tempo, key and mood arc first, then the voice, then the instruments and how they change from section to section. MiniMax recommends exactly that three-part structure.
