Five minutes of generated music with MiniMax Music 3

MiniMax Music 3 generates songs up to five minutes long with control over lyrics, arrangement, and musical structure.

MiniMax Music 3 generates complete songs up to five minutes long from lyrics and a detailed music description. The model is designed to maintain themes, rhythm, vocal identity, and the progression of the arrangement over time, with structures that can include intros, verses, choruses, bridges, instrumental sections, and outros.

Control is based on two complementary inputs. Lyrics can include tags such as `[Verse]`, `[Chorus]`, `[Bridge]`, or `[Solo]`, while a separate description defines the genre, BPM, key, emotional progression, vocal characteristics, instrumentation, and how those elements evolve throughout the track. MiniMax also provides a tool that can turn a short instruction into a more detailed Structured Caption.

The architecture separates the song’s overall coherence from its acoustic details. An 8B-parameter Global LLM, initialized from Qwen3-8B, handles long-range musical progression. A 600M-parameter Local LLM then adds acoustic information at the frame level. The internal states of both models are fused before passing through a 2.4B-parameter Flow Matching stage and then a Flow-VAE responsible for reconstructing the audio signal.

This architecture avoids relying exclusively on the tokenizer’s discrete tokens during final synthesis. MiniMax says the continuous representations produced by the LLMs preserve more information related to vocal articulation, instrumental textures, and temporal continuity. The output is generated as 32 kHz, 16-bit stereo WAV audio.

The model can be downloaded and run with SGLang-Omni, Diffusers, or ComfyUI. The Diffusers documentation indicates that full-precision inference can fit on a GPU with less than 24 GB of VRAM, using around 22 GB with CPU offloading. A slower mode can also reduce requirements to as little as 8 GB of VRAM by progressively loading the language model layers.

Local execution still comes with several limitations. Generation currently works only in non-streaming mode, the text prompt is limited to 5,000 tokens, and output is capped at 9,000 acoustic frames. MiniMax also notes that tempo, key, instrumentation, lyrics, and structural instructions act as generative controls rather than strict guarantees, meaning the result may deviate from some requested details.