MiniMax Releases MiniMax-Music3: An Open-Weights Music Model Generating Complete Five-Minute Songs From Lyrics and a Structured Caption
MiniMax launched MiniMax-Music3, an open-weights text-to-music mannequin. The mannequin takes two separate inputs: lyrics carrying part tags, and a detailed music description. It returns a full tune of as much as 5 minutes in a single technology, as 32 kHz, 16-bit stereo WAV. The structure pairs a Hybrid-LM, an 8B Global LLM with a 0.6B Local LLM, with a steady synthesis stack constructed on movement matching and a Flow-VAE. Weights, inference code and three documented serving paths shipped the identical day.
Is it deployable?
Yes, MiniMax printed usable weights, inference code and three documented serving paths on day one, so that is deployable now slightly than a analysis preview.
- Company degree: Solo creators, indie studios and mid-market groups can ship on it immediately. The MiniMax-Music3 Community License permits industrial use, nevertheless it requires you to show ‘MiniMax-Music3’ prominently within the product UI, and any group whose mixture yearly income from these merchandise exceeds US$ 20 million should receive separate prior written authorization from MiniMax. Anyone internet hosting third-party technology should additionally implement and preserve safeguards in opposition to infringing outputs.
- Industries: Game growth, promoting and model businesses, short-form video and creator instruments, e-learning, podcasting, health and wellness apps, retail in-store audio, and music-tech SaaS.
- Applications: Background scoring for UGC video, adaptive recreation and degree music, localized advert beds and sonic branding, scratch and demo tracks for songwriters, mood-conditioned playlist technology, and offline batch technology the place per-song API value is the constraint.
The Architecture
MiniMax-Music3 combines a hierarchical autoregressive stack with a steady synthesis path.
The coaching tokenizer makes use of eight layers of residual vector quantization (RVQ). The first, semantic codebook has 16,384 entries and carries core musical semantics and construction. The remaining seven acoustic codebooks have 1,024 entries every and encode residual element. Training optimizes the semantic layer first, then all eight collectively.
The Hybrid-LM splits the modeling downside. An 8B Global LLM predicts the primary RVQ codebook body by body and holds long-range construction; a 0.6B Local LLM predicts the remaining codebooks inside every body. The mannequin card and license state the Global LLM is initialized from Qwen3-8B; the MiniMax Research post says Qwen3.5-8B, so deal with the precise base checkpoint as unsettled.
The synthesis stage is the extra fascinating design selection. Rather than decoding from discrete RVQ tokens, MiniMax fuses the ultimate hidden states of each LLMs and situations a 2.4B flow-matching module on them, which maps into a latent area decoded by a 123M Flow-VAE inherited from MiniMax Speech. At inference the discrete tokenizer decoder shouldn’t be loaded in any respect.
Two-input management
Lyrics carry the phrases and part tags on their very own strains: [Intro], [Verse], [Pre-Chorus], [Chorus], [Post-Chorus], [Bridge], [Instrumental], [Solo], [Outro]. A separate Structured Caption carries Global Metadata, Vocal Details and Arrangement. MiniMax additionally ships a music-caption-rewriter agent ability that expands a brief description into that three-part format offline.