MiniMax Music 3.0

AI model · Music & audio generation · Text → Audio / speech by MiniMax

Next-Generation Open-Weights, Production-Ready & Versatile Music Model

MiniMax Music 3.0 is a next-generation music generation model designed for creators and developers to produce production-ready songs. Given a creative concept and optional lyrics, the model independently handles composition, arrangement, performance, and production in a single generation. It utilizes a global-local Hybrid-LM architecture to sustain expressive intent and structural stability for tracks up to five minutes long. The model is released with open weights, featuring an 8B Global LLM and a 0.6B Local LLM to deliver high-fidelity vocals and realistic instrumental rendering.

Languages
English, Mandarin
Global LLM Parameters
8B
Local LLM Parameters
0.6B
Flow-Matching Parameters
2.4B
Flow-VAE Parameters
123M
Maximum Song Duration
5 minutes
Architecture
Hybrid-LM with RVQ and Flow-VAE

Overview

MiniMax Music 3.0 is a next-generation music generation model designed for creators and developers to produce production-ready songs. Given a creative concept and optional lyrics, the model independently handles composition, arrangement, performance, and production in a single generation. It utilizes a global-local Hybrid-LM architecture to sustain expressive intent and structural stability for tracks up to five minutes long. The model is released with open weights, featuring an 8B Global LLM and a 0.6B Local LLM to deliver high-fidelity vocals and realistic instrumental rendering.

Music Creation

  • 5-Minute Duration: Music 3.0 can sustain expressive intent and structural stability across a complete song of up to five minutes.
  • High-Fidelity Rendering: The model renders instruments with physical realism and generates vocals that sound performed rather than synthesized.

Technical Architecture

  • Hierarchical RVQ Tokenizer: An eight-layer residual vector quantization (RVQ) represents musical information hierarchically, separating core structure from acoustic residuals.
  • Flow-Based Synthesis Stack: A 2.4B flow-matching module and a 123M Flow-VAE decode continuous hidden states into high-fidelity audio.

Specifications

Global LLM Parameters
8B
Local LLM Parameters
0.6B
Flow-Matching Parameters
2.4B
Flow-VAE Parameters
123M
Maximum Song Duration
5 minutes
Architecture
Hybrid-LM with RVQ and Flow-VAE
License
Open Weights

Added by AntSpace from public sources (minimax.io).

Related

More AI models by MiniMax