MiniMax Music 3.0 turns a style prompt and your lyrics into a finished stereo MP3 song with vocals and backing. Leave the lyrics empty to have them written, or choose instrumental; songs ran about one to four minutes in tests.
{
"model": "minimax/music-3.0/text-to-audio",
"input": {
"generate_lyrics": true,
"instrumental": false,
"bitrate": 128000,
"sample_rate": 44100,
"lyrics": "[Verse]\nHeadlights painting gold along the shore\nYour hand keeps drifting closer than before\nThe radio is playing something slow\nI don't want to know which way is home\n\n[Pre Chorus]\nEvery red light is a reason to stay\n\n[Chorus]\nSo drive, drive, keep the city in the mirror\nStay, stay, let the night pull us nearer\nWe've got nowhere to be and all of the time\nMidnight looks so good in your eyes\n\n[Outro]\nMidnight looks so good in your eyes",
"prompt": "Dreamy synth-pop, 100 bpm, airy female lead vocal, shimmering pads, pulsing bassline, gated drums, late-night city drive mood"
}
}
Introduction
MiniMax Music 3.0 on SpicyAPI is a text-to-music task that returns one complete song per request: a lead vocal, backing arrangement and final mix in a single stereo MP3. You send a style brief in prompt and, if you have them, the words to sing in lyrics; the model handles melody, arrangement and performance. Songs came back between 1:09 and 3:58 in our tests, and every song costs the same flat price.
Key features & highlights
A whole song, not a loop
The result is a structured song with an intro, verses, a chorus that comes back and an ending, rather than a short repeating clip. The model sets the length itself, mostly from how many lyrics you give it, and mixes vocal and backing into one finished stereo file that is ready to play.
Three ways to handle the words
You can sing your own lyrics, have the model write them, or leave the singer out. Lyrics you provide are performed as written, and section tags such as [Verse] and [Chorus] steer the arrangement. With lyrics empty, generate_lyrics (on by default) writes a full set of lyrics from the prompt. With instrumental on, the track has no vocals at all.
Flat price, adjustable file quality
Length, bitrate and sample_rate do not change what a song costs. Pick 256 kbps at 44.1 kHz for a master, or a lower bitrate for small preview files.
Request parameters
Request parameters for MiniMax Music 3.0 · Text to Music
Parameter
Required
Description
prompt
Yes
The style brief, up to 2,000 characters: genre, mood, tempo, instruments, the kind of voice, the setting. at least 1 characters · at most 2000 characters
lyrics
No
The words to sing, up to 3,000 characters. Put each line on its own row and leave an empty row for a pause. at least 1 characters · at most 3000 characters
generate_lyrics
No
On by default: when lyrics is empty, the model writes the lyrics from your prompt. Lyrics you provide are left untouched. Default: true
instrumental
No
Switch on for music with no vocals; no lyrics are needed. Off by default. Default: false
bitrate
No
MP3 bitrate in bits per second: 32000, 64000, 128000 or 256000 (default). Lower values make smaller files. 32000 · 64000 · 128000 · 256000 · Default: 256000
sample_rate
No
Sample rate in Hz: 16000, 24000, 32000 or 44100 (default). Does not change the price. 16000 · 24000 · 32000 · 44100 · Default: 44100
Pricing
$0.18 / song
One flat price per song, whatever its length or audio quality.
Use cases
Songs built around your own lyrics
Write the lyric sheet first, tag the sections, and describe the genre and voice you hear in your head. This suits personalised songs, character themes for stories and games, parody tracks and rough demos of a lyric you want to hear sung before taking it further.
Music for video and streams
Turn on instrumental for underscores, lounge beds and intro music, or keep the vocal for a title song. A one-to-four-minute track covers most short videos in a single request, and the flat price makes it cheap to generate a few takes and keep the best one.
Use prompt for the sound only: genre, tempo in bpm, two or three instruments, the kind of voice and the mood, for example “dreamy synth-pop, 100 bpm, airy female lead vocal, shimmering pads, late-night city mood”. Put everything that should be sung in lyrics. Style words left in the lyrics tend to get sung.
Shape the song with section tags
Write one lyric line per row and put tags such as [Intro], [Verse], [Pre Chorus], [Chorus], [Bridge] and [Outro] on their own rows. Repeat the chorus text where you want it to return. A verse and a chorus give a song of about a minute and a half; add verses and a bridge for a longer one.
Know what the filter stops
MiniMax checks prompts and lyrics. Strong language is sung as written, but lyrics beyond that are refused and are not charged; keep those lines mild.
API quick start
# Generate the idempotency key once; reuse it when retrying after a lost response.
export SPICY_IDEMPOTENCY_KEY="${SPICY_IDEMPOTENCY_KEY:-$(uuidgen)}"
curl -X POST https://api.spicyapi.ai/api/v1/jobs/createTask \
-H "Authorization: Bearer $SPICY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $SPICY_IDEMPOTENCY_KEY" \
-d '{
"model": "minimax/music-3.0/text-to-audio",
"input": {
"prompt": "Calm ambient pads with a slow rising motif",
"bitrate": 256000,
"sample_rate": 44100
}
}'
# Copy data.taskId from the response into TASK_ID, then poll the status.
TASK_ID="TASK_ID_FROM_RESPONSE"
curl "https://api.spicyapi.ai/api/v1/jobs/recordInfo?taskId=$TASK_ID" \
-H "Authorization: Bearer $SPICY_API_KEY"
See the API reference for authentication, the complete schema, polling, and error handling.
Related models and next tasks
Missing the model you need?
Send a style brief and lyrics, get a finished song with vocals
MiniMax Music 3.0 is MiniMax's song model, released on 2026-07-16, with open weights also published. One request returns one complete song as a stereo MP3: a lead vocal, a backing arrangement and a finished mix. You describe the style in prompt and, if you have them, send words in lyrics. The model decides how long the song runs, so there is no duration field; in our tests songs ran from 1:09 to 3:58. Clean lyrics are sung exactly as written and strong language comes through, but MiniMax's own filter refuses anything past that. Billing is one flat price per song, shown live on this page.
One flat price per song; refused requests are not charged
Scope
A whole song from one requestYou do not assemble stems or stitch clips together. The response is a single MP3 with a lead vocal, the band behind it and a final mix, so a lyric sheet becomes something you can play to a client or drop on a timeline straight away. Length is the model's decision, which suits songs built around words rather than a stopwatch. A verse and a chorus came out near a minute and a half, and a full song with auto-written lyrics ran close to four minutes.
Prompting
Style words in prompt, singable words in lyricsThe prompt field takes a style brief of 1 to 2,000 characters and the lyrics field takes up to 3,000 characters, one line per row, with a blank row for a pause. Section tags on their own rows, such as Verse, Chorus, Bridge and Outro in square brackets, shape the arrangement. Keep genre, tempo and mood out of the lyrics, because anything left in lyrics gets sung along with the rest.
Tested
What our runs showed on 2026-10-05Every clean lyric line was sung as written. A duet prompt with no lyrics returned a 3:58 song with two voices trading lines. Turning instrumental on gave a 2-minute jazz lounge track with no vocal. A punk chorus full of swearing was sung word for word. Lyrics from our content test were refused after about 90 seconds and not charged, and the strongest ones never finished and timed out.
What the MiniMax Music 3.0 API does well
Lyrics sung as you wrote them
Send lyrics and the model sings those words rather than paraphrasing them. In our test every clean line came back as written. That makes it useful for checking whether a lyric sheet scans, whether a chorus lands and where a line needs another syllable, before anyone books a studio or a singer.
Lyrics written for you
generate_lyrics is on by default. If lyrics is empty, the model writes words from your prompt and then sings them. This is the slowest path, since auto-written lyrics take longest, but it turns a one-line idea into a full song. Lyrics you send are never replaced, they are sung as written.
Duets and trading voices
Describe two singers in the prompt and the model can split the song between them. Our duet prompt, sent with no lyrics at all, returned a 3:58 song with two voices trading lines. Name who sings what in the prompt, such as a playful male and female pair, to steer the exchange.
Instrumental mode
Switch instrumental on and you get a track with no vocal and no lyrics needed. Our smoky jazz lounge prompt returned a 2-minute track with brushed drums, bass, muted trumpet and electric piano in the arrangement and no singing. It is the clean route to menu loops, vlog beds and background music.
Arrangement control with section tags
Put a tag on its own row and the arrangement follows it. The supported set is Intro, Verse, Pre Chorus, Chorus, Post Chorus, Hook, Bridge, Interlude, Transition, Break, Build Up, Inst, Solo and Outro, each in square brackets. Use Inst or Solo rows to open space for the band.
Audio quality you can dial
bitrate accepts 32, 64, 128 or 256 kbps with 256 as the default, and sample_rate accepts 16, 24, 32 or 44.1 kHz with 44.1 as the default. Drop them for quick drafts or small files, keep the defaults for finished work. The price per song stays the same whatever you pick.
Who MiniMax Music 3.0 is for
Songwriters and lyricists
Hear the lyric sheet before booking a session
Paste a verse and a chorus with section tags, add a style brief and listen to a sung version in a few minutes. Because lyrics are sung as written, you learn quickly where a line is crowded, where the hook lands and whether the mood matches your idea, and you can rewrite before spending studio time.
Video creators
A theme song or vlog soundtrack in one request
Describe the feel of the channel or the piece, add two lines of lyrics or let the model write them, and receive a full song with a lead vocal. For wordless beds under talking, turn instrumental on. Either way you get one MP3 ready to place on the timeline.
Game and app developers
Background songs and menu loops
Set instrumental on and write the mood into the prompt: tempo, instruments, setting. You get a finished track with no vocal to compete with dialogue or interface sounds. Run several briefs for different screens or levels and compare them, since each request returns a complete piece.
Gift and personal songs
A birthday or anniversary song from a few lines of story
Write the story in the prompt, such as names, a shared memory and the mood you want, leave lyrics empty and keep generate_lyrics on. The model writes the words and sings them, usually landing around the four-minute mark. Keep the themes warm and clean, as MiniMax refuses anything past strong language.
How to call the MiniMax Music 3.0 API
01
Create an API key
Sign in to SpicyAPI, create a key and keep it on your server. Every request carries it as a bearer token, and you can also test the same inputs in the playground first to hear what the model does with your brief.
02
Write the style brief
Put genre, tempo, mood, vocal type and instruments into prompt, up to 2,000 characters. Be specific, for example dreamy synth-pop at 100 bpm with an airy female lead. Leave story and singable words for the lyrics field.
03
Add lyrics or let the model write them
Send lyrics up to 3,000 characters, one line per row, a blank row for a pause and section tags on their own rows. Or leave lyrics empty and keep generate_lyrics on. For a track with no vocal, set instrumental to true.
04
Set quality and submit
Optionally choose bitrate and sample_rate, otherwise the defaults of 256 kbps and 44.1 kHz apply. Submit the task. There is no duration_seconds here because the model picks the length itself.
05
Wait, then download the MP3
Generation takes about 1.5 to 4 minutes, and auto-written lyrics take longest. When the task finishes you receive a stereo MP3 link. Download it, because generated audio is kept for 14 days. Failed or refused requests are not charged.
MiniMax Music 3.0 worked examples
Dreamy synth-pop from a short lyric sheet
A style brief in prompt and a verse and chorus in lyrics, each with its section tag. Expect the airy vocal to sing the lines as written over pads and a pulsing bass, and a song of roughly a minute and a half.
{"prompt": "Dreamy synth-pop, 100 bpm, airy female lead vocal, shimmering pads, pulsing bassline, late-night city drive mood", "lyrics": "[Verse]\nHeadlights painting gold along the shore\nYour hand keeps drifting closer than before\n\n[Chorus]\nSo drive, drive, keep the city in the mirror\nMidnight looks so good in your eyes"}
A duet written from the prompt alone
No lyrics are sent, so generate_lyrics writes them. The story and the two voices are described in the prompt. A similar duet brief in our tests returned a 3:58 song with two voices trading lines.
{"prompt": "Upbeat Latin pop duet about a summer fling that starts on a crowded beach and ends at sunrise, playful male and female vocals trading lines, nylon guitar, 96 bpm"}
A jazz lounge instrumental
Instrumental is switched on, so no lyrics are needed and no vocal appears. Our run returned a 2-minute track, a good template for background music and menu loops.
{"prompt": "Smoky late-night jazz lounge instrumental, brushed drums, upright bass, muted trumpet and Rhodes piano, slow 72 bpm", "instrumental": true}
MiniMax Music 3.0 API questions
It is an API for MiniMax's song model, released on 2026-07-16. You send a style brief in prompt, optionally lyrics, and one request returns one complete stereo MP3 song with a lead vocal, a backing arrangement and a final mix. The model chooses the length. SpicyAPI exposes it with the same task flow as its other media models.
Billing is one flat price per song, regardless of length, bitrate or sample rate, so changing those settings does not change the cost. The live price is shown on this page before you run anything. Failed or refused requests are not charged, which includes lyrics that MiniMax declines to sing.
No. This is a filtered model. Strong language is sung, and our punk chorus full of swearing came back word for word, but lyrics from our content test were refused after about 90 seconds and not charged, and the strongest ones never finished and timed out.
The model decides, and there is no duration field. Our songs ran from 1:09 to 3:58: a verse and a chorus came out near a minute and a half, and a full song with auto-written lyrics lasted about four minutes. For an instrumental bed at a set length under a video, use Seed Audio 1.0, which takes the length in its prompt.
Yes. Send up to 3,000 characters in lyrics, one line per row, with a blank row for a pause and section tags such as Verse or Chorus on their own rows. Lyrics you send are sung as written. If you leave lyrics empty, generate_lyrics, which is on by default, writes them from your prompt.
Yes. Set instrumental to true and you get a track with no vocal, with no lyrics needed. Our jazz lounge prompt returned a 2-minute instrumental. Describe tempo, instruments and mood in prompt, and use it for background songs, menu loops and video beds.
When you want the model to write and sing a whole song from clean lyrics without deciding a length. For an instrumental bed under a video, use Seed Audio 1.0 or "Score this cut" in the Studio video editor instead.
MiniMax made it and released it on 2026-07-16, and open weights have also been published. SpicyAPI offers it through one API and one account next to other music models. Generated audio is kept for 14 days, so download your songs before then.
Treat the two text fields as two jobs. Put everything about sound in prompt: genre, tempo in bpm, mood, the type of voice and a few instruments, for example dreamy synth-pop at 100 bpm with an airy female lead and shimmering pads. Put only words that should be sung in lyrics, one line per row, with a blank row where you want a pause. Style words left in lyrics get sung, so keep them out. Mark the structure with section tags on their own rows, such as Verse, Pre Chorus, Chorus, Bridge and Outro in square brackets, and add Inst or Solo rows when the band should take over. Short, singable lines work better than long sentences. Keep the lyrics clean, because MiniMax refuses anything past ordinary strong language, which is sung. Start with a verse and a chorus to hear the voice and the mix, then extend. If you only have an idea, leave lyrics empty and let generate_lyrics write the song, accepting a longer wait.
All MiniMax models21
Audio models3
MiniMax Music 3.0CurrentText to audio$0.18per request