Description
MiniMax Music 3 is a sophisticated music generation model designed to create complete songs with a duration of up to five minutes. This model is conditioned on both lyrics and a detailed music description, allowing it to generate structurally coherent songs with expressive vocals, evolving arrangements, and stable long-form audio quality. The model combines an 8B Global LLM for long-range musical structure and a 0.6B Local LLM for frame-level acoustic detail, utilizing a continuous hidden-state synthesis system based on Flow Matching and Flow-VAE. It produces 32 kHz, 16-bit stereo WAV audio.
Read More
The model supports full-song generation, maintaining musical themes, rhythm, vocal identity, and arrangement progression across long sequences. It enables complete structures such as intro, verse, pre-chorus, chorus, bridge, instrumental break, and outro. Users can input lyrics with explicit section tags and a music description defining the musical style, emotional progression, vocal performance, instrumentation, arrangement, and production profile.
MiniMax Music 3 employs a hierarchical autoregressive architecture that separates global musical modeling from local acoustic modeling. The Global LLM predicts the first RVQ codebook frame by frame, modeling the song's long-range semantic and structural progression. The Local LLM predicts the remaining acoustic codebooks within each frame, restoring fine-grained acoustic information.
The model's synthesis module fuses the final hidden states of the Global and Local LLMs, preserving richer acoustic information for vocal articulation, instrumental texture, and temporal continuity. The Flow-VAE architecture is adapted from MiniMax Speech and retrained for the dynamic range and spectral characteristics of music.
MiniMax Music 3 is supported by SGLang-Omni and can be served using various inference frameworks such as SGLang, diffusers, and ComfyUI. It requires CUDA for inference and currently supports only non-streaming generation. The tokenized text prompt is limited to 5,000 tokens, and audio generation is limited to 9,000 acoustic frames.
- CategoryAI Music Generators
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsWeb
MiniMax Music 3 Highlights
High-performance music generation
Supports full-song generation up to five minutes
Combines Global and Local LLMs
Produces 32 kHz, 16-bit stereo WAV audio
Hierarchical autoregressive architecture
Continuous hidden-state synthesis
Flow-VAE architecture
Supports SGLang-Omni and other frameworks
Getting Started with MiniMax Music 3
Access model: Visit Hugging Face
Authenticate: Set up account
Set up environment: Follow installation guide
Integrate via API: Use supported frameworks
Optimise: Adjust parameters for desired output
MiniMax Music 3's Use Cases
- Songwriting
- Music Production
- Film Scoring
- Game Soundtracks
- Advertising
