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-weighted music generation model. The model employs a layered architecture of 8B Global LLM and 0.6B Local LLM, combined with Flow Matching + Flow-VAE continuous hidden-state synthesis technology. It can generate complete 32kHz stereo songs up to 5 minutes long based on lyrics and structured musical descriptions. The model supports fine-grained segment control, maintaining consistency in theme, rhythm, and vocal identity throughout long audio tracks, achieving end-to-end generation from creative intent to a high-fidelity final product.

MiniMax Music 3.0's main features

• Full song generationIt supports generating complete songs up to 5 minutes long, covering standard structures such as intro, verse, chorus, bridge, and outro.
• Lyrics-driven creationEnter the lyrics and a description of the music style to generate a complete work that includes vocals and instrumental arrangement.
• Structured paragraph control:support [Verse]、[Chorus]、[Bridge] Tags such as these allow for precise control over the mood and arrangement changes in each section of the song.
• Long-range consistency maintenanceMaintaining a stable musical theme, rhythm, vocal identity, and arrangement progression throughout an audio recording lasting several minutes.
• High-fidelity audio outputGenerates 32kHz, 16-bit stereo WAV files with sound quality approaching professional recording studio standards.
• Automatic enhancement of prompt words:built-in music-caption-rewriter The tool can expand a brief description into a professional, structured prompt that includes global metadata, vocal details, and arrangement planning.

The technical principles of MiniMax Music 3.0

• Hierarchical Autoregressive Architecture (Hybrid-LM)The model consists of an 8B Global LLM and a 0.6B Local LLM working together. The Global LLM is initialized based on Qwen3 and predicts the first RVQ semantic codebook frame by frame, responsible for modeling the long-range structure and semantics of the song; the Local LLM predicts the remaining acoustic codebooks in each frame, recovering fine-grained timbre and sound quality. The two are jointly trained to achieve efficient separation and collaboration between macroscopic structure and microscopic acoustics.
• Multilevel residual vector quantization (RVQ)An 8-layer RVQ is used to represent musical information hierarchically. The first layer semantic codebook contains 16,384 entries, capturing core musical semantics and structure; the remaining 7 layers acoustic codebooks each contain 1,024 entries, encoding residual acoustic details level by level. During training, the semantic codebooks are first optimized independently to build a stable skeleton, and then all codebooks are trained jointly to balance semantic capacity, generation stability, and detail reconstruction.
• Continuous hidden state synthesisThe final continuous hidden states of the Global and Local LLMs are fused through inference and used as conditional input. The 2.4B Flow Matching module maps these states to the VAE latent space, and then the 123M Flow-VAE decodes them into waveforms. This process preserves richer high-dimensional acoustic information, significantly improving the accuracy of vocal pronunciation, the physical realism of instruments, and long-term consistency.
• Structured description frameworkIt introduces a structured description system with fine time granularity, expanding a single global tag into a three-part description that includes Global Metadata (genre, BPM, tonality), Vocal Details (timbre, singing style, harmony), and Arrangement (instrument evolution, section changes), so that subjective creative intentions are transformed into professional arrangement instructions that the model can execute.

Follow us on WeChat and reply with "open source",join inAI open source project discussion group

How to use MiniMax Music 3.0

• Environmental preparationEnsure you have two NVIDIA GPUs that support CUDA on your local machine, and configure the Python environment and the SGLang-Omni inference framework dependencies.
• Download Model:implement hf download MiniMaxAI/MiniMax-Music3 --local-dir /path/to/minimax_ttm The command pulls the complete model weights from HuggingFace to the local directory.
• Start service:run sgl-omni serve --model-path MiniMaxAI/MiniMax-Music3 --port 8000 Start the inference service, where GPU 0 is responsible for Qwen3 and eight-layer RVQ autoregressive generation, and GPU 1 is responsible for Flow Matching and Flow-VAE waveform decoding.
• Ready to inputIn API requests input Enter paragraph labels in the field (e.g.) [Verse]、[Chorus]The lyrics of ) instructions The fields are filled with music style, mood, and arrangement description, which together serve as the generation conditions.
• Send Request:Towards http://127.0.0.1:8000/v1/audio/speech Send a POST request and set model for minimax_ttm、response_format for wav、max_new_tokens Setting it to 9000 will return the generated 32kHz stereo WAV audio.
• Enhanced prompts (optional)Install and call music-caption-rewriter The skill automatically expands a brief music description into a structured prompt that includes global metadata, vocal details, and section arrangement, thereby improving the accuracy and controllability of the generation.

MiniMax Music 3.0's core advantages

• Full long song generationIt natively supports the generation of complete songs up to 5 minutes long, maintaining consistency in musical theme, rhythm, vocal identity, and arrangement progression in long audio files, achieving a truly "complete song".
• Accurately understand the creative intentIt transforms subjective creative ideas into executable professional arrangement instructions through a structured caption framework, preventing the generated content from deviating from the original requirements, and provides prompts and enhancement tools to lower the professional threshold.
• High-fidelity audio renderingIt employs continuous hidden state synthesis (blending LLM hidden states → Flow Matching → Flow-VAE) instead of traditional discrete token decoding, outputting 32kHz stereo sound, resulting in a more realistic physical feel for instruments and a more transparent mix.
• Precise and controllable arrangementIt supports lyric segment tags and a three-part structured description (global metadata, vocal details, and segment arrangement), and allows for precise timing control of emotional evolution, instrument entry and exit, and vocal techniques.
• Natural vocal performanceSignificantly reduces high-frequency digital artifacts, improves pronunciation accuracy, breathiness, and harmonic layering, making the generated vocals closer to real singing than synthetic effects.

MiniMax Music 3.0 project address

• 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 competing products

Application scenarios of MiniMax Music 3.0

• Quick Demo Creation for Independent MusiciansMusicians can input original lyrics and style descriptions, and generate a complete demo including vocals and arrangement within minutes, greatly reducing the trial and error costs in the early stages of creation.
• Customized soundtracks for short videos and podcastsContent creators can generate theme songs or background music based on the emotional curve of videos, realizing the transformation from "finding copyrighted music" to "customizing exclusive music".
• Dynamic soundtracks for games and interactive entertainmentDevelopers integrate the model into the game engine, dynamically generating long-duration music that adapts to the scene atmosphere based on the player's real-time behavior, thereby enhancing the sense of immersion.
• Audio content for advertising and brand marketingThe advertising team can quickly produce multiple versions of advertising jingles and songs based on the brand's tone and style, supporting A/B testing and avoiding copyright disputes.
• Music Education and Composition Teaching SupportTeachers use models to demonstrate the actual effects of different genres, structures, and arrangement techniques, allowing students to intuitively understand the principles of music production by adjusting structured descriptions.