MiniMax Music 3.0 - MiniMax's Open-Source Music Generation Model

What is MiniMax Music 3.0

MiniMax Music 3.0 is MiniMax's next-generation open-weightmusic generationmodel. The model adopts a hierarchical architecture of an 8B Global LLM and a 0.6B Local LLM, combined with Flow Matching + Flow-VAE continuous latent state synthesis technology, and can generate complete songs up to 5 minutes long in 32kHz stereo based on lyrics and structured music descriptions. The model supports fine-grained section control, maintaining consistency of theme, rhythm, and vocal identity across long audio, achieving end-to-end generation from creative intent to high-fidelity finished product.

Main Features of MiniMax Music 3.0

• Complete Song Generation: Supports generating complete songs up to 5 minutes long, covering standard structures such as intro, verse, chorus, bridge, and outro.
• Lyric-Driven Creation: Input lyrics and a music style description to generate a complete work including vocal performance and instrumental arrangement.
• Structured Section Control: Supports [Verse]、[Chorus]、[Bridge] and other tags to precisely control the mood and arrangement changes of each section of the song.
• Long-Range Consistency Maintenance: Stably maintains the musical theme, rhythm, vocal identity, and arrangement progression across audio lasting several minutes.
• High-Fidelity Audio Output: Generates 32kHz, 16-bit stereo WAV files, with sound quality approaching professional studio standards.
• Automatic Prompt Enhancement: Built-in music-caption-rewriter tool, which can expand a short description into a professional structured prompt containing global metadata, vocal details, and arrangement planning.

Technical Principles of MiniMax Music 3.0

• Hierarchical Autoregressive Architecture (Hybrid-LM): The model works through the collaboration of an 8B Global LLM and a 0.6B Local LLM. The Global LLM is initialized based on Qwen3 and predicts the first RVQ semantic codebook frame by frame, responsible for modeling the song's long-range structure and semantics; the Local LLM predicts the remaining acoustic codebooks within each frame, restoring fine-grained timbre and sound quality. Through joint training, the two achieve efficient separation and collaboration between macro structure and micro acoustics.
• Multi-Layer Residual Vector Quantization (RVQ): Uses 8-layer RVQ for hierarchical representation of music information. The first-layer semantic codebook contains 16,384 entries, capturing core musical semantics and structure; the remaining 7 layers of acoustic codebooks each contain 1,024 entries, encoding residual acoustic details progressively. During training, the semantic codebook is first optimized independently to establish a stable skeleton, and then all codebooks are jointly trained to balance semantic capacity, generation stability, and detail reconstruction.
• Continuous Latent State Synthesis: Inference fuses the final continuous latent states of the Global and Local LLMs, which are used as conditional input to a 2.4B Flow Matching module mapped to the VAE latent space, and then decoded into waveforms by a 123M Flow-VAE. This preserves richer high-dimensional acoustic information, significantly improving vocal pronunciation accuracy, instrumental physical realism, and long-term consistency.
• Structured Description Framework: Introduces a structured description system with fine temporal granularity, expanding a single global tag into a three-part description comprising Global Metadata (genre, BPM, key), Vocal Details (timbre, singing style, harmony), and Arrangement (instrument evolution, section changes), transforming subjective creative intent into professional arrangement instructions executable by the model.

Follow WeChat and reply "open source" to join theAI Open-Source Project Exchange Group

How to use MiniMax Music 3.0

• Environment preparation: Ensure you have two CUDA-capable NVIDIA GPUs locally, and set up the Python environment as well as the dependencies of the SGLang-Omni inference framework.
• Download the model: Run the hf download MiniMaxAI/MiniMax-Music3 --local-dir /path/to/minimax_ttm command to pull the complete model weights from HuggingFace to a local directory.
• Start the service: Run sgl-omni serve --model-path MiniMaxAI/MiniMax-Music3 --port 8000 to start the inference service, where GPU 0 handles Qwen3 and the eight-layer RVQ autoregressive generation, and GPU 1 handles Flow Matching and Flow-VAE waveform decoding.
• Prepare the input: In the API request's input field, enter the lyrics with paragraph tags (such as [Verse]、[Chorus]), and in the instructions field, enter the music style, mood, and arrangement description; together they serve as the generation conditions.
• Send the request: Send a POST request to http://127.0.0.1:8000/v1/audio/speech , set model to minimax_ttm、response_format to wav、max_new_tokens to 9000, and the generated 32kHz stereo WAV audio will be returned.
• Prompt enhancement (optional): Install and invoke the music-caption-rewriter skill to automatically expand a brief music description into a structured prompt containing global metadata, vocal details, and section arrangement, in order to improve generation precision and controllability.

Core advantages of MiniMax Music 3.0

• Complete long-song generation: Native support for generating complete songs up to 5 minutes long, stably maintaining consistency of the musical theme, rhythm, singing voice identity, and arrangement progression throughout the long audio, achieving a true "complete song."
• Precise understanding of creative intent: Through the Structured Caption framework, subjective creativity is transformed into executable professional arrangement instructions, avoiding generated content that deviates from the original requirements, and equipped with prompt enhancement tools to lower the professional threshold.
• High-fidelity audio rendering: Uses continuous latent state synthesis (fusing LLM latent states → Flow Matching → Flow-VAE), rather than traditional discrete token decoding, outputting 32kHz stereo, with stronger physical realism of instruments and clearer mixing.
• Finely controllable arrangement: Supports lyric paragraph tags and three-part structured description (global metadata, vocal details, section arrangement), enabling precise timing control over emotional progression, instrument entrances and exits, and vocal techniques.
• Natural vocal performance: Significantly reduces high-frequency digital artifacts, improves pronunciation accuracy, breathiness, and harmony layering, making the generated vocals closer to real singing rather than a synthetic effect.

Project addresses of MiniMax Music 3.0

• Project official website：https://www.minimax.io/blog/minimax-music-3-0-next-generation-open-weights-production-ready-versatile-music-model
• GitHub repository：https://github.com/MiniMax-AI/MiniMax-Music3

Comparison of MiniMax Music 3.0 with similar competitors

Application scenarios of MiniMax Music 3.0

• Rapid Demo production for independent musicians: Musicians input original lyrics and style descriptions to generate a complete Demo including vocals and arrangement within minutes, greatly reducing the trial-and-error cost in the early stage of creation.
• Customized soundtrack for short videos and podcasts: Content creators generate theme songs or background music according to the emotional curve of the video, realizing the transformation from "searching for copyrighted music" to "customizing exclusive music".
• Dynamic soundtrack for games and interactive entertainment: Developers integrate the model into game engines to dynamically generate long-duration music that adapts to the scene atmosphere based on players' real-time behavior, enhancing immersion.
• Audio content for advertising and brand marketing: Advertising teams quickly produce multiple versions of advertising songs and jingles based on brand tone descriptions, supporting A/B testing and avoiding copyright disputes.
• Music education and creative teaching assistance: Teachers use the model to demonstrate the actual effects of different genres, structures, and arrangement techniques, and students intuitively understand the principles of music production by adjusting structured descriptions.