Independent guide · Not affiliated with MiniMax
MiniMax Music 3.0
MiniMax Music 3.0 is the common search name for MiniMax Music 3, MiniMax’s official open-weight model for generating complete songs from lyrics and detailed musical direction.
This independent guide explains the documented five-minute song limit, model architecture, GPU paths, Structured Caption format, license conditions, and current difference between the open weights and the hosted MiniMax music API.
MiniMax Music 3.0 key facts
primary sources only
- Last verified
- Official model name
- MiniMax Music 3
- Release channel
- Public model weights
MiniMax published the model card, model repository, community license, and a first-party audio demo under its official organizations. The model card documents songs up to five minutes and 32 kHz, 16-bit stereo WAV output. The official title omits “.0”; this page uses MiniMax Music 3.0 only as a search-friendly alias.
Documented in the model card
What MiniMax Music 3 officially supports
These are published capabilities of the open-weight Music 3 release. They are not benchmark claims from this site: we have not run or independently evaluated the model, and we preserve MiniMax’s stated limits.
Songs up to five minutes
The official model card says MiniMax Music 3 can generate complete songs up to five minutes while maintaining themes, rhythm, vocal identity, and arrangement progression.
Lyrics plus detailed direction
Lyrics provide the words and section tags; a separate music description controls genre, mood, vocals, instrumentation, arrangement, and production profile.
Structured song control
Supported lyric tags include Intro, Verse, Pre-Chorus, Chorus, Post-Chorus, Bridge, Instrumental, Solo, and Outro.
Local, open-weight deployment
MiniMax publishes model weights and setup paths for SGLang-Omni, Diffusers, and ComfyUI rather than limiting Music 3 to a hosted product interface.
Source: official MiniMax Music 3 model card.
Architecture in numbers
Inside the Music 3 stack
MiniMax describes a hierarchical autoregressive system that separates long-range structure from local acoustic detail, then synthesizes audio from fused continuous hidden states.
- 8B
Global LLM
Models long-range musical semantics and structure through the first RVQ codebook.
- 0.6B
Local LLM
Restores frame-level acoustic information through the remaining codebooks.
- 2.4B
Flow Matching module
Turns fused hidden states into the latent representation used for waveform synthesis.
- 123M
Flow-VAE decoder
Decodes the music latent into final stereo audio.
- 32 kHz
Sample rate
The documented output is 32 kHz, 16-bit stereo WAV audio.
- 8 layers
RVQ tokenizer
One 16,384-entry semantic codebook plus seven 1,024-entry acoustic codebooks are used during training.
Architecture source: MiniMax Music 3 model card. Parameter counts describe components; they are not independent quality scores.
Local tool · runs in your browser
MiniMax Music 3 Structured Caption Builder
The official model card recommends three description sections: Global Metadata, Vocal Details, and Arrangement. Choose a few parameters to draft those sections locally. Lyrics remain a separate model input; this tool does not generate audio, call MiniMax, or upload your choices.
Global Metadata Genre: Indie folk. Tempo: Mid (90–110 BPM). Key and scale: C major. Emotional progression: begin intimate, build toward a hopeful chorus, and resolve with a deliberate outro. Listening scenario: A late-night drive through an empty city. Production profile: detailed stereo image, controlled low end, clear transients, and section-to-section dynamic contrast. Vocal Details Single lead vocal, clearly mixed above the instrumentation. Performance: natural phrasing with a distinct hook; keep diction clear and backing parts subordinate to the lead. Arrangement Use instrumentation native to Indie folk; introduce layers gradually, change the groove at structural transitions, and reserve the fullest arrangement for the final chorus. Structure: Intro → Verse → Pre-Chorus → Chorus → Verse → Chorus → Bridge or Instrumental → Final Chorus → Outro.
Generated locally from your selections. Add your own lyrics separately, review every musical detail, and remember that the model card describes tempo, key, instrumentation, and structure as generative controls rather than strict guarantees.
Do not mix the delivery paths
Open-weight Music 3 vs. the hosted music API
MiniMax Music 3 is an official model release, but the public hosted API reference checked on 2026-08-14 still documents Music 2.6 identifiers. Treat self-hosted Music 3 and the hosted API as separate access paths unless MiniMax updates that reference.
| Topic | Music 3 open weights | Hosted music API docs |
|---|---|---|
| Distribution | Public model weights | Hosted MiniMax API |
| Documented identifier | MiniMaxAI/MiniMax-Music3 | music-2.6 / music-2.6-free |
| Primary access path | SGLang-Omni, Diffusers, or ComfyUI | POST /v1/music_generation |
| Infrastructure | Your CUDA-capable machine | MiniMax-managed infrastructure |
| Output documented in source | 32 kHz, 16-bit stereo WAV | Configurable API audio settings |
| Music 3 hosted API ID | Not listed in the public hosted API docs checked | Public docs currently list Music 2.6 models |
Sources: Music 3 model card and hosted music-generation API reference.
Before you download
Requirements and limitations
- Inference requires CUDA.
- Only non-streaming generation is currently supported.
- The tokenized text prompt is limited to 5,000 tokens.
- Audio generation is limited to 9,000 acoustic frames.
- Tempo, key, instruments, lyrics, and structure are generative controls, not strict symbolic guarantees.
The official card also says full precision fits under 24 GB of VRAM, CPU offloading uses about 22 GB, and layer streaming can fit an 8 GB GPU more slowly. Treat those as published setup guidance, not a benchmark from this site.
Community license
Commercial-use notes
- The community license covers software, model weights, parameters, configuration, inference code, and documentation.
- Commercial products using the model must prominently display “MiniMax-Music3” in their interface.
- Products above the license’s stated US$20 million aggregate annual-revenue threshold require prior written authorization from MiniMax.
- Hosted generation services must implement and maintain safeguards against prohibited or rights-infringing uses.
This is a plain-language reading aid, not legal advice. The license and Acceptable Use Policy control.
Read the official licenseFrequently asked
Questions about MiniMax Music 3.0
Hear the model, then read the source
Start with MiniMax’s audio examples, then use the model card for architecture, setup, limitations, and links to supported runtimes. This independent page does not host the weights or run inference.