MiniMax Music 3.0 API by MiniMax

MiniMax Music 3.0

Developer: MiniMax Model ID: minimax/music-3.0 Release Date: August 13, 2026

Overview

MiniMax Music 3.0 is MiniMax's next-generation, open-weights music model. It takes a musical description and (optionally) lyrics and returns a complete, finished song of up to five minutes in a single generation — composed, arranged, performed, and produced. There is no separate melody step, no stem assembly, and no post-production pass: one request yields a mixed track with lead vocals, harmonies, and full instrumental backing.

The model's design goal is expressive fidelity over the length of a whole song. Where earlier text-to-music systems tend to drift — losing the mood, the groove, or the vocal identity somewhere after the first chorus — Music 3.0 is built to understand a creator's intent and then sustain that intent from intro to outro, with instruments that sound physically played and vocals that sound performed rather than synthesized.

The Music 3.0 Model

Music 3.0 is an approximately 11.1B-parameter system assembled from four specialized components rather than one monolithic network. Each stage handles a different scale of musical information.

Architecture

Hierarchical audio tokenizer (8-layer RVQ). Music is represented with eight layers of residual vector quantization. The first layer, backed by a 16,384-entry codebook, carries core semantics and musical structure; the remaining seven layers, each with a 1,024-entry codebook, progressively encode acoustic residuals. This separation is what lets the model reason about song-level form independently of fine timbral detail.

Hybrid global–local language model. An 8B global LLM (initialized from Qwen3.5-8B) predicts semantic tokens frame by frame and is responsible for long-range structure — form, emotional arc, harmonic movement. A 0.6B local LLM predicts the acoustic tokens within each frame, capturing residual detail. The two are trained in stages: global alignment first, then joint parameter training.

Continuous synthesis path. A 2.4B flow-matching module consumes fused hidden states from both language models and produces VAE hidden states, which a 123M Flow-VAE decoder (inherited from MiniMax's speech architecture and retrained for music) renders into audio. Because the synthesis path consumes continuous hidden states rather than discrete tokens, far richer high-dimensional acoustic information survives into the final waveform.

Structured Captions and Prompt Enhancement

Music 3.0 was trained on Structured Captions — annotations with fine temporal granularity that describe emotional contour, the entry and exit of individual instruments, groove development, changes in vocal delivery, harmony, and vocal effects across the timeline of a track. This is why the model responds to production-level direction instead of only genre labels.

A built-in prompt enhancement system expands a short, casual description into detailed musical instructions, so a brief prompt still produces a coherent, fully specified arrangement.

What Improved in 3.0

• Semantic understanding — the upgraded semantic model markedly reduces the structural drift typical of AI-generated songs, holding a single creative intent across all five minutes.
• Instrumental realism — arrangements are more complete and varied, with fine-grained temporal control over when instruments enter and leave, and reproduction of authentic performance techniques such as glissando and legato.
• Mix quality — mixes are more open, clear, and balanced, with much less congestion and muddiness, approaching commercial-recording-grade listening.
• Vocal performance — the new vocal engine removes high-frequency digital hiss and improves pronunciation, breathing, phrasing, and harmony control. Harmonies unfold more naturally and breathing becomes part of the performance rather than an artifact.

Key Capabilities

• Full songs in one pass — up to five minutes of finished music per request, including vocals, arrangement, and mix.
• Section-level structural control — inline structure tags placed on their own line steer the arrangement: [Intro], [Verse], [Pre Chorus], [Chorus], [Post Chorus], [Hook], [Bridge], [Interlude], [Transition], [Build Up], [Break], [Inst], [Solo], [Outro].
• Vocal or instrumental — a single flag switches between a sung track and a purely instrumental piece.
• Automatic lyric writing — the model can write its own lyrics directly from the musical description, so no separate lyric-writing step is required.
• Production-level prompt control — prompts can specify genre, mood, scenario, tempo, key, instrumentation, vocal timbre and delivery, and production character (for example "deep bouncy 808 sub-bass, rolling triplet hi-hats, male vocalist with heavy autotune, cynical nocturnal energy").
• Broad genre and multilingual range — demonstrated across Mandarin pop, EDM, bossa nova, funk, pop-rock, indie, industrial, and baroque pop, with both Chinese and English vocals.
• Configurable audio delivery — choice of MP3, WAV, or PCM, with selectable sample rate and (for MP3) bitrate.
• Open weights — the model weights are publicly released under the MiniMax-Music3 Community License, which permits commercial deployment with visible attribution; organizations whose aggregate yearly revenue from such products exceeds US$20 million require separate prior written authorization from MiniMax.

Input

There are two creative inputs, and they play different roles:

• prompt describes the music — style, mood, tempo, instrumentation, vocal character, production.
• lyrics supplies the words — line-separated text with optional section tags controlling the arrangement.

How you combine them selects one of three modes:

1. Lyrics-driven song — supply lyrics (with or without a prompt).
2. Prompt-driven song — supply a prompt, set lyrics_optimizer to true, and leave lyrics empty; the model writes the lyrics itself.
3. Instrumental — set is_instrumental to true and supply a prompt; lyrics are ignored.

Every request must satisfy at least one of these three shapes.

Input Parameters

Parameter	Type	Default	Range / Options	Description
model	string	minimax/music-3.0	—	Required. Model name.
prompt	string	upbeat electronic pop, bright female vocal, catchy hook, energetic, 120bpm	≤ 2000 characters	Musical description of the song: genre, mood, scenario, tempo, key, instrumentation, vocal timbre and delivery, and production character. Required when is_instrumental is true, and when lyrics is empty and lyrics_optimizer is true.
is_instrumental	boolean	false	—	Generate instrumental music only (no vocals). When true, lyrics is ignored and only prompt is used.
lyrics_optimizer	boolean	false	—	Let the model write the lyrics from the prompt. Requires a non-empty prompt and an empty lyrics field; it cannot be combined with lyrics you supply yourself. Only takes effect when is_instrumental is false.
lyrics	string	""	≤ 3500 characters	Song lyrics. Separate lines with a newline; use a blank line for a pause. Supports section structure tags on their own line (e.g. [Verse], [Chorus]) to control the arrangement of the song. Required unless is_instrumental is true, or lyrics_optimizer is true and prompt is non-empty — in which case the model writes the lyrics from the prompt and this field must be left empty.
format	string	mp3	mp3, wav, pcm	Output audio format.
sample_rate	integer	44100	16000, 24000, 32000, 44100	Output sample rate (Hz).
bitrate	integer	256000	32000, 64000, 128000, 256000	Output bitrate (bps). Applies to the mp3 format only.
enable_sync_mode	boolean	false	—	Wait for the result to be generated and uploaded before returning, so the audio is delivered directly in the response (API only).
enable_base64_output	boolean	false	—	Return the output as a Base64 string instead of a URL (API only).

Parameter Constraints

The schema enforces three conditional rules, so it is worth stating them plainly:

• If is_instrumental is true, prompt is required and must be non-empty.
• If lyrics_optimizer is true, lyrics must be empty — the two are mutually exclusive.
• At least one of the following must hold: is_instrumental is true; lyrics is non-empty; or prompt is non-empty and lyrics_optimizer is true.

Output

Returns the generated song as an audio file — by default a URL in the outputs array, or a Base64-encoded string when enable_base64_output is enabled. Requests are asynchronous by default: the response carries a request ID that can be polled until the track is ready, and enabling sync mode returns the finished audio directly.

The underlying model renders stereo audio natively; format, sample_rate, and bitrate control how that audio is encoded for delivery. Alongside the audio, the response reports task status, executionTime, an inference timing breakdown, and NSFW detection flags per output.

Use Cases

• Content soundtracks — original background music and theme songs for video, podcasts, and streams, cleared of third-party licensing concerns.
• Advertising and brand audio — jingles, campaign tracks, and sonic branding tailored to a brief in a single request.
• Games and interactive media — instrumental loops, level themes, and mood beds generated from production-level prompts.
• Songwriting and demos — turning a lyric sheet into a fully arranged, fully mixed reference recording for pitching or iteration.
• Personalized music — greetings, dedications, and occasion songs generated at scale from a short description.
• Music apps and agentic pipelines — programmatic song generation inside a product or workflow, with structure tags giving deterministic control over arrangement.

Pricing

Pricing for Music 3.0 is a flat rate per generated song. It does not vary with prompt length, lyric length, song duration, output format, sample rate, or bitrate — a thirty-second instrumental sketch and a full five-minute vocal track cost exactly the same. Enabling lyrics_optimizer so the model writes its own lyrics is also included at no extra charge.

SKUs

SKU	Description	Unit Price
sku_base	Per generated song (flat, up to 5 min)	$0.15

Formula

cost = sku_base

The formula consists of the single base SKU, so every successful generation is billed $0.15, and total spend is simply the number of songs generated multiplied by that rate.

Examples

Scenario	Calculation	Cost
Song from supplied lyrics	0.15	$0.1500
Song with model-written lyrics (lyrics_optimizer)	0.15	$0.1500
Instrumental track (is_instrumental)	0.15	$0.1500
Full 5-minute WAV at 44.1 kHz	0.15	$0.1500
10 songs	0.15 × 10	$1.5000

Because the rate is flat, cost forecasting is exact: budget $0.15 for every generation you intend to run, regardless of how long or how richly produced the resulting song is.

Related Models

minimax/lyrics-generation is the natural companion to this model. It writes structured lyrics — complete with section tags and style tags — from a theme, and its output is directly usable as the lyrics input here. Use it when you want to review or hand-edit lyrics before committing to a full generation; use this model's lyrics_optimizer flag when you would rather go from concept to finished song in one call.