MiniMax Music 3.0 – MiniMax's Open-Weight Music Generation Model

1. What is MiniMax Music 3.0

MiniMax Music 3.0 is the next-generation open-weight music generation model introduced by MiniMax. The model employs a hierarchical architecture combining an 8B Global LLM and a 0.6B Local LLM, and integrates Flow Matching and Flow-VAE continuous latent state synthesis technologies. It can generate complete songs up to 5 minutes long in 32kHz stereo audio based on lyrics and structured music descriptions. It supports fine-grained control over song segments, maintaining consistency in theme, rhythm, and vocal identity throughout long audio tracks, enabling end-to-end generation from creative intent to high-fidelity output.

Image source: Official article
Image source: official article

Technical Positioning and Domain: MiniMax Music 3.0 belongs to the AI music generation domain, focusing on end-to-end generation of complete songs and precise structural control. Unlike existing models that mainly target short clips or pure instrumental generation, this model is positioned as a production-grade, locally deployable open-weight solution, aiming to lower the barrier to entry for professional music creation while delivering audio quality and control capabilities comparable to cloud-based services.

Development Background: Developed by the MiniMax team, which has deep expertise in multimodal large models (such as the MiniMax-01 series) and audio generation technologies, Music 3.0 was launched in response to the shortcomings of current music generation models in terms of long-term consistency, control granularity, and audio quality. The goal is to provide a solution that meets professional needs while maintaining an open ecosystem, allowing creators to truly use AI as a collaborative tool rather than just a toy.

Core Value: It addresses the common issues in current music generation models, such as "being able to generate clips but struggling to complete full songs" and "being able to mimic styles but unable to precisely control structure." Through a hierarchical architecture and a structured description framework, creators can precisely drive composition using lyrics and segment labels, maintaining thematic and vocal identity consistency across up to 5 minutes of audio. This significantly enhances the practicality and artistic completeness of generated content, thereby shifting AI music generation from "experimental" to "usable."

Technical Features: The core technology lies in the dual-engine design of "hierarchical autoregressive + continuous latent state synthesis." The 8B Global LLM handles macro-level semantics and structure, while the 0.6B Local LLM focuses on micro-level acoustic details, working together to achieve efficient generation. During the inference phase, Flow Matching is introduced to map the LLM latent states into the VAE latent space, and then Flow-VAE decodes them into waveforms, preserving richer acoustic information. This significantly improves the accuracy of vocal articulation and the physical realism of instruments. The structured description framework enables subjective creativity to be translated into executable composition instructions for machines.

2. Key Features

• Full Song Generation: Supports the generation of complete songs up to 5 minutes in length, covering standard structures such as intro, verse, chorus, bridge, and outro. Unlike most models that can only generate snippets of 30 seconds to 1 minute, this feature enables creators to obtain complete musical works with a full narrative arc in one go, reducing the workload of post-production assembly and truly achieving "a complete song."

• Lyric-Driven Composition: By inputting lyrics and a music style description, the model can generate a complete piece that includes vocal performance and instrumental arrangement. Lyrics support structured paragraph tags (e.g., [Verse], [Chorus]), allowing the model to precisely align lyric sections with musical sections, achieving a high degree of synchronization between lyrics and music, and avoiding issues of misalignment between lyrics and melody.

• Structured Paragraph Control: Using tags such as [Verse], [Chorus], [Bridge], creators can precisely control the emotion, arrangement density, and instrument entry/exit for each section. This capability stems from the model's training on a structured description framework, enabling it to understand and execute detailed instructions like "add strings to the second chorus," offering a level of arrangement controllability far beyond that of natural language description models.

• Long-Term Consistency Maintenance: Over several minutes of audio, the model can consistently maintain the coherence of musical themes, rhythm, vocal identity, and arrangement progression. This is made possible by the Global LLM's modeling of the overall song structure and the Local LLM's coherent generation of details, preventing common issues in long audio such as thematic drift or sudden tonal shifts, and ensuring a unified style throughout the entire piece.

• High-Fidelity Audio Output: Generates 32kHz, 16-bit stereo WAV files with audio quality approaching professional recording studio standards. By using continuous hidden state synthesis instead of traditional discrete token decoding, it minimizes quantization loss, resulting in more authentic instrument tones and clearer mixing, especially excelling in high-frequency vocal details and breathiness, significantly reducing the "digital" or "electronic" sound.

• Prompt Auto-Enhancement: Built-in music-caption-rewriter tool automatically expands brief descriptions (e.g., "a sad pop song") into professional structured prompts that include global metadata, vocal details, and arrangement planning. This lowers the usage barrier for non-expert users while providing advanced users with a more precise starting point, enhancing generation accuracy and controllability.

3. How to Use

1. Environment Requirements and Preparation: Requires at least two NVIDIA GPUs that support CUDA (recommended: A100 40GB or equivalent computational power). The system must have Python 3.8+ installed, along with the SGLang-Omni inference framework and its dependencies (such as torch, transformers, etc.). The VRAM requirement depends on parameters such as max_new_tokens, and it is recommended that the VRAM per GPU be no less than 24GB. For optimal compatibility, it is advised to use Linux (Ubuntu 20.04+).

2. Download Model Weights: Execute the command hf download MiniMaxAI/MiniMax-Music3 --local-dir /path/to/minimax_ttm to pull the full model weights from HuggingFace. The total model size is approximately 8B + 0.6B parameters, corresponding to tens of gigabytes. Ensure that sufficient disk space is available (it is recommended to reserve more than 100GB). The download process may take a long time, and a tool for resuming interrupted downloads can be used.

3. Start Inference Service: Run sgl-omni serve --model-path MiniMaxAI/MiniMax-Music3 --port 8000. The service will automatically assign GPU 0 to handle Qwen3 and the eight-layer RVQ autoregressive generation, and GPU 1 to handle Flow Matching and Flow-VAE waveform decoding. You can specify the GPU order by setting the environment variable CUDA_VISIBLE_DEVICES, or adjust --max-running-requests to control the concurrency level.

4. Construct Request and Generate: Send a POST request to http://127.0.0.1:8000/v1/audio/speech. The request body must include: set model to minimax_ttm, fill the input field with lyrics containing paragraph tags (e.g., [Verse]First verse lyrics[/Verse]), fill the instructions field with a description of the music style and arrangement (e.g., "Pop, 120BPM, C major, using piano and drums"), set response_format to wav, and set max_new_tokens to 9000 (approximately corresponding to 5 minutes of audio). The returned WAV file can be played directly or saved.

5. Prompt Enhancement (Optional): Install and invoke the music-caption-rewriter skill to automatically expand a brief description (e.g., "A relaxing jazz song") into a structured prompt that includes global metadata, vocal details, and arrangement for each paragraph. Then pass this structured prompt into the instructions field, which can significantly improve generation accuracy. This tool can be installed via pip or directly obtained from GitHub.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

Selection Recommendations: For professional musicians who require full song generation, fine-grained control, and are willing to invest in hardware costs, MiniMax Music 3.0 is currently the preferred choice among open-weight solutions. Its structured description framework and long-term consistency capabilities stand out among similar tools, making it especially suitable for scenarios requiring repeated iteration on composition details. For users who prioritize the highest audio quality and longest duration and are not concerned about cloud dependency, Suno v5.5 offers superior output specifications (48kHz 24-bit) and duration (8 minutes). However, it is fully closed-source and cannot be customized locally. For creators who need to generate short audio clips quickly and emphasize iteration speed, Udio performs well in terms of generation speed and quality for short audio, but lacks the level of control provided by MiniMax and has limited duration. For scenarios requiring complete offline operation and only short music clips (e.g., game sound effects or alert tones), MusicGen is an open-source model that is easy to deploy and has low hardware requirements. However, it significantly lags behind MiniMax in terms of functionality and audio quality, making it unsuitable for full song generation.

6. Editor's Summary

MiniMax Music 3.0 demonstrates significant innovation in its technical architecture. Its design, combining "hierarchical autoregressive + continuous hidden state synthesis," separates macro-level structural modeling from micro-level acoustic details, ensuring long-term consistency while enhancing audio quality. This approach is more reasonable compared to traditional single autoregressive or diffusion models and holds reference value both academically and in engineering. The introduction of a structured description framework further translates subjective creativity into executable instructions, greatly improving controllability—a capability currently lacking in music generation models. It fills the gap in the open-weight domain regarding "complete song generation."

In terms of practical value, this model enables musicians to deploy it locally, avoiding cloud latency and copyright concerns, while allowing for rapid iteration of demos through precise control, thereby reducing production costs. For content creators and game developers, the ability to generate customized music is highly appealing, especially for scenarios requiring frequent adjustments and batch generation. Although the hardware requirements are relatively high, for studio and enterprise users, the one-time hardware investment offers unlimited local generation, making long-term costs manageable.

The target users mainly include: independent musicians needing to quickly produce demos, video and podcast creators seeking efficient background music, game developers aiming to dynamically generate ambient tracks, and researchers interested in AI music technology. For general enthusiasts, cloud services may be more convenient, but MiniMax's open ecosystem provides a stronger foundation for technical exploration.

Looking ahead, as community contributions grow and documentation improves, MiniMax Music 3.0 has the potential to become a key foundation in open-source music generation. Future support for more languages, extended generation duration, and reduced hardware requirements (such as the release of quantized or distilled versions) could further expand its application boundaries. Overall, this is a work that achieves new heights in both technical depth and practical value, setting a new benchmark for open-weight models in the AI music generation field.

7. Application Scenarios

• Rapid Demo Production for Independent Musicians: Musicians input original lyrics and style descriptions, and the model generates a complete demo with vocals and instrumentation within minutes. It supports repeated adjustments to the structure of sections (e.g., modifying the density of the chorus instrumentation), significantly reducing the trial-and-error cost of composition and allowing the creative focus to return to the lyrics and melody itself. This is especially suitable for small studios with limited budgets.

• Customized Music for Short Videos and Podcasts: Content creators write structured descriptions based on the emotional curve of the video (e.g., "soft piano for the first 30 seconds, then add drum beats to reach a climax after 30 seconds"), generating background music that synchronizes with the visual rhythm, achieving "customized exclusive music." This avoids copyright risks while enhancing content originality and brand recognition.

• Dynamic Music for Games and Interactive Entertainment: Developers integrate the model into game engines to dynamically generate long-duration music that matches the scene based on real-time parameters (e.g., battle status, exploration atmosphere), enhancing immersion. For example, when players enter different areas, the model switches instrumentation in real time according to predefined style tags, reducing the workload of traditional music production.

• Audio Content for Advertising and Brand Marketing: Advertising teams quickly generate multiple versions of jingles and ad songs based on brand tone descriptions (e.g., "youthful energy, electronic style, with vocal slogans"), supporting A/B testing. This avoids legal risks associated with copyrighted music and improves production efficiency, adapting to the fast-paced iteration of marketing campaigns.

• Music Education and Composition Instruction: Teachers use the model to demonstrate the practical effects of different genres, structures, and composition techniques. Students can intuitively understand the principles of music production by adjusting structured descriptions (e.g., changing BPM or instrument configurations), making it an interactive teaching tool that reduces the abstraction of theoretical learning.

8. FAQ

Q: Does MiniMax Music 3.0 support Chinese lyrics?
A: The model is primarily optimized for English lyrics, but in theory, it can accept Chinese lyrics as input. However, due to the predominantly English training data, the pronunciation accuracy and naturalness of Chinese lyrics may not be as high as that of English lyrics, potentially resulting in mispronunciations or unnatural intonation. We recommend users test it themselves, and in the future, support for more languages may be added through fine-tuning or community contributions.

Q: How much VRAM is required for deployment?
A: It is recommended to use two NVIDIA A100 (40GB or 80GB) GPUs. If using consumer-grade GPUs (e.g., RTX 4090 24GB), you may need to adjust the batch size during service startup (e.g., --max-running-requests 1) or use a quantized version (if available from the official source). The exact VRAM usage depends on parameters such as max_new_tokens, and generating a 5-minute audio file may require more than 20GB of VRAM on a single GPU.

Q: Can it generate pure instrumental music (without vocals)?
A: Yes. Leave the lyrics input blank or only input section tags (e.g., [Instrumental]), and describe the instrumental style in the instructions. The model can then generate pure instrumental pieces. However, the model is mainly designed for vocal song generation, and control over pure instrumental music may not be as refined as with specialized models. It is recommended to explicitly specify the instrument combination in your description.

Q: Compared to Suno v5, which one is better?
A: They serve different purposes. MiniMax Music 3.0 is an open-weight model that can be deployed locally, offering finer control and suitability for scenarios requiring customization and data privacy. Suno v5 is a closed-source cloud service with higher audio quality (44.1kHz/48kHz), longer duration (v5.5 supports up to 8 minutes), and no hardware investment required. The choice depends on specific needs: if control and data security are priorities, choose MiniMax; if ease of use and highest audio quality are desired, choose Suno.

Q: Who owns the copyright to the generated works?
A: Since the model uses an open-weight license (check the official LICENSE file for details), the copyright of the output content generated by users is typically determined by the user themselves. However, note that if the input used during model operation is protected by copyright (such as existing melodies or lyrics), there may be a risk of infringement. We recommend reviewing MiniMax's license terms and ensuring that the input content is original or has been authorized.

Q: Is commercial use supported?
A: According to MiniMax's open-weight license, commercial use is generally allowed, but the license terms (such as Apache 2.0 or custom licenses) must be followed. For specific details, please refer to the LICENSE file in the GitHub repository, which will clearly state whether commercial use is permitted and if attribution is required. We recommend reading it carefully before commercial deployment.

9. Project Links

• Product 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
• SGLang-Omni Framework: https://github.com/sgl-project/sglang (used for deploying inference services)
• HuggingFace Model Page: https://huggingface.co/MiniMaxAI/MiniMax-Music3