MiniMax Music 3 is a high-performance music generation model for creating complete songs up to five minutes long. Conditioned on lyrics and a detailed music description, it generates structurally coherent songs with expressive vocals, evolving arrangements, and stable long-form audio quality.
MiniMax Music 3 combines an 8B Global LLM for long-range musical structure, a 0.6B Local LLM for frame-level acoustic detail, and a continuous hidden-state synthesis system based on Flow Matching and Flow-VAE. The model produces 32 kHz, 16-bit stereo WAV audio.
MiniMax Music 3 natively supports full-song generation up to five minutes. The model maintains musical themes, rhythm, vocal identity, and arrangement progression across long sequences, enabling complete structures such as intro, verse, pre-chorus, chorus, bridge, instrumental break, and outro.
Fine-Grained Music Control
The model accepts two complementary inputs:
Lyrics define the words to be sung and may include explicit section tags such as [Intro], [Verse], [Pre-Chorus], [Chorus], [Post-Chorus], [Bridge], [Instrumental], [Solo], and [Outro].
Music description defines the musical style, emotional progression, vocal performance, instrumentation, arrangement, and production profile.
For precise control, we recommend using a Structured Caption with three sections:
Global Metadata: genre, subgenre, BPM, key, scale, emotional progression, listening scenario, and production profile.
Arrangement: primary and secondary instruments, section-level instrument evolution, groove, bass, percussion, textures, and spatial effects.
This representation allows the model to follow not only a global style, but also the musical development of the song over time.
Hybrid-LM
MiniMax Music 3 uses a hierarchical autoregressive architecture that separates global musical modeling from local acoustic modeling.
The Global LLM (8B) predicts the first RVQ codebook frame by frame and models the song's long-range semantic and structural progression.
The Local LLM (0.6B) predicts the remaining acoustic codebooks within each frame and restores fine-grained acoustic information.
The Global LLM is initialized from Qwen3-8B. During training, its embedding and output layers are first adapted to semantic music tokens. The Global and Local LLMs are then jointly trained to model all RVQ codebooks.
Continuous Hidden-State Synthesis
Instead of decoding only from discrete RVQ tokens, the synthesis module fuses the final hidden states of the Global and Local LLMs. These continuous representations preserve richer acoustic information for vocal articulation, instrumental texture, and temporal continuity.
The Flow-VAE architecture is adapted from MiniMax Speech and retrained for the dynamic range and spectral characteristics of music.
Music Tokenizer
The training tokenizer uses eight layers of Residual Vector Quantization (RVQ):
The first semantic codebook contains 16,384 entries and captures the core musical semantics and structure.
The remaining seven acoustic codebooks contain 1,024 entries each and represent residual acoustic details.
Training first optimizes the semantic codebook, then jointly trains all eight codebooks. At inference time, waveform synthesis uses the fused LLM hidden states and does not require the discrete tokenizer decoder.
How to Use
MiniMax Music 3 is supported by SGLang-Omni. Follow the official installation guide to prepare the runtime environment.
The service uses the shared speech API. Put the lyrics in input and the music description in instructions. Put lyric structure tags such as [Verse] and [Chorus] on their own lines.
bash
1curl http://127.0.0.1:8000/v1/audio/speech \2 -H 'Content-Type: application/json'\3 -d '{
4 "model": "MiniMaxAI/MiniMax-Music3",
5 "input": "[Verse]\nMorning light filtering through the pine\n[Chorus]\nSoftly the world begins to breathe",
6 "instructions": "A warm acoustic pop song with intimate female vocals, fingerpicked guitar, soft piano, and a gradual emotional build into a wide final chorus.",
7 "response_format": "wav",
8 "seed": 7,
9 "max_new_tokens": 750,
10 "stream": false
11 }'\12 --output minimax_music3.wav
max_new_tokens sets the maximum number of audio frames at 25 frames per second. Generation may finish before this limit when the model emits an end-of-audio token. The response is a 32 kHz, 16-bit stereo WAV file.
Reproducible Example
The following end-to-end example contains the complete lyrics, music description, and generation parameters used to produce the reference audio.
1import soundfile as sf
2import torch
3from diffusers import ModularPipeline
45pipe = ModularPipeline.from_pretrained("MiniMaxAI/MiniMax-Music3")6pipe.load_components(dtype=torch.bfloat16)7pipe.to("cuda")89lyrics ="""[verse]
10Morning light filtering through the pine
11Every quiet street is yours and mine
12[chorus]
13Softly the world begins to breathe"""1415prompt =(16"Genre: acoustic pop. BPM: 96. Key: C major. Warm and intimate, building gently into the chorus. "17"Vocals: soft female lead, close and breathy, light stacked harmonies in the chorus. "18"Arrangement: fingerpicked guitar and soft piano; brushed drums and upright bass enter in the chorus."19)2021audio = pipe(22 prompt=prompt,23 lyrics=lyrics,24 audio_duration=60.0,25 generator=torch.Generator("cuda").manual_seed(7),26 output="audios",27)[0]2829sf.write("song.wav", audio.T.float().cpu().numpy(), pipe.sampling_rate)
Low VRAM
The full precision fits under 24GB of VRAM. With automatic CPU offloading, generation takes in ~22 GB; additionally streaming the language model layer by layer makes it fit even 8 GB video cards:
python
1import torch
2from diffusers import ComponentsManager, ModularPipeline
3from diffusers.hooks import apply_group_offloading
45manager = ComponentsManager()6manager.enable_auto_cpu_offload(device="cuda")7pipe = ModularPipeline.from_pretrained("MiniMaxAI/MiniMax-Music3", components_manager=manager)8pipe.load_components(dtype=torch.bfloat16)910# Only needed below ~22 GB of VRAM — slower, but fits in 8 GB.11apply_group_offloading(12 pipe.language_model, onload_device=torch.device("cuda"), offload_type="leaf_level", use_stream=True13)1415
Prompt Enhancement
A concise natural-language description can be used directly. For richer prompts and more precise control, use the provided music-caption-rewriter skill to expand it into a Structured Caption containing Global Metadata, Vocal Details, and Arrangement. The skill preserves musical instructions attached to lyric section tags in the arrangement description while keeping the lyric text in the lyrics input.
Only non-streaming generation is currently supported.
The tokenized text prompt is limited to 5,000 tokens.
Audio generation is limited to 9,000 acoustic frames.
Section tags and music descriptions provide generative control rather than strict symbolic guarantees. The generated tempo, key, instrumentation, lyrics, and song structure may not always match every requested detail exactly.
