MiniMax Open-Sources Weights for Music 3 AI Music Generation Model

Chinese company MiniMax has released the weights and code for its MiniMax Music 3 AI music generation model. The model can generate complete audio tracks with vocals and instrumentation up to 5 minutes long, with its files and a web demo available on Hugging Face.

The model synthesizes stereo tracks in 32kHz, 16-bit WAV format while preserving traditional song structures with verses, choruses, bridges, and solos. Generation is controlled via two input fields: lyrics and music description. The developers recommend dividing the music description into three parts: style metadata (genre, tempo in BPM, key), vocal parameters, and instrument arrangement details.

Architecture and System Requirements

Music 3 is built on a two-tier architecture:

• Global LLM (8 billion parameters): based on Qwen3-8B, responsible for overall song composition and progression over time;
• Local LLM (0.6 billion parameters): generates acoustic details and vocal transitions;
• Flow Matching and Flow-VAE: handle the final audio waveform synthesis.

The full repository size is 57.4GB. For standard operation, the developer recommends a graphics card with 24GB of VRAM, though CPU offloading support allows generation to run on consumer GPUs with 8GB of VRAM. The model supports integration with the Diffusers and SGLang frameworks, as well as the ComfyUI node-based environment.

License Terms

The model is distributed under the MiniMax-Music3 Community License. It can be used for free for local testing and commercial services, provided the company’s annual revenue does not exceed $20 million.

Files and model weights are available in the official MiniMax repository on Hugging Face, while technical deployment instructions are provided in the platform documentation.