MiniMax Music 3: Long-Form AI Music Generation Model

The artificial intelligence music generation ecosystem is reaching a major milestone with the arrival of MiniMax Music 3. This high-performing model stands out for its ability to generate complete tracks lasting up to five minutes, while maintaining structural coherence and high acoustic quality in 32 kHz, 16-bit stereo.\n\nThe industry is increasingly moving toward hybrid architectures capable of handling both global semantics and local acoustic details. MiniMax Music 3 perfectly illustrates this trend by combining large language models with advanced synthesis techniques like Flow Matching.\n\n## Hybrid Architecture and Hierarchical Modeling\n\nTo successfully generate music over such long durations without losing the melodic or rhythmic thread, MiniMax Music 3 relies on an Hybrid-LM type architecture. This approach separates the modeling of global musical structure from frame-level acoustic details.\n\n- The Global LLM (8B), initialized from Qwen3-8B, predicts the first RVQ (Residual Vector Quantization) codebook frames image by image. It manages semantic progression and long-term structure (intro, verse, chorus, bridge, etc.).\n- The Local LLM (0.6B) takes over to predict the remaining acoustic codebooks within each frame, thereby restoring the finesse and texture of the sound.\n\nRather than decoding solely from discrete tokens, the system fuses the final hidden states of both LLMs. These continuous representations preserve crucial information for vocal articulation, temporal continuity, and instrumental textures.\n\n## Tokenization and RVQ Representation\n\nThe model's audio tokenizer relies on eight layers of residual vector quantization (RVQ). The first (semantic) codebook captures overall structure and melody, while codebooks 2 through 8 (acoustic) reproduce the details and coloration of the signal.\n\n## Fine Control and Prompt Structuring\n\nMiniMax Music 3 accepts two complementary input streams to guide generation: lyrics and musical description.\n\nTo effectively structure a track, it is recommended to use explicit section tags directly within the lyrics, such as [Intro], [Verse], [Pre-Chorus], [Chorus], [Bridge], [Solo], and [Outro].\n\nMusical description can be enriched via a structured caption divided into three axes:\n1. Global metadata: genre, sub-genre, tempo (BPM), key, scale, emotional progression, and production profile.\n2. Vocal details: gender of the voice, timbre, performance style, harmonies, and vocal effects.\n3. Arrangement: primary and secondary instruments, instrumental evolution per section, groove, bass, percussion, and spatial effects.\n\n## Implementation and Serving with SGLang-Omni\n\nMiniMax Music 3 is fully integrated into the open-source ecosystem, notably via SGLang-Omni and Hugging Face's diffusers library. Inference strictly requires a CUDA-compatible environment and supports non-streaming generation. Lyric text is limited to 5,000 tokens, and audio generation is capped at 9,000 acoustic frames.