MiniMax has released Music 3.0, an open-weight music generation model that produces complete, structurally coherent songs up to five minutes long from a text description and optional lyrics. The model is available on Hugging Face and integrated into ComfyUI, with API access through MiniMax's platform.
The system combines an 8B Global LLM initialized from Qwen3-8B for long-range musical structure with a 0.6B Local LLM for frame-level acoustic detail. Instead of decoding audio from discrete tokens alone, a continuous synthesis pipeline fuses hidden states from both language models through a 2.4B Flow Matching module and a 123M Flow-VAE decoder, outputting 32 kHz, 16-bit stereo WAV audio.
What's new
- Full-song generation up to five minutes with native support for complete structures: intro, verse, pre-chorus, chorus, bridge, instrumental break, and outro.
- Lyrics with explicit section tags — [Intro], [Verse], [Pre-Chorus], [Chorus], [Post-Chorus], [Bridge], [Instrumental], [Solo], [Outro] — giving creators control over song architecture.
- Structured Captions for fine-grained control across three sections: Global Metadata (genre, BPM, key, emotional progression, production profile), Vocal Details (gender, timbre, performance style, harmony, effects), and Arrangement (instruments, section-level evolution, groove, bass, percussion, spatial effects).
- Multi-layer RVQ tokenizer with eight codebooks: a 16,384-entry semantic layer for core structure and seven 1,024-entry acoustic layers for residual detail.
- Hybrid-LM architecture with two-stage training: global alignment first, then joint training of Global and Local LLMs.
- Hidden-state fusion synthesis bypassing discrete token decoding for richer vocal articulation and instrumental texture.
- Open weights under MiniMaxAI/MiniMax-Music3 on Hugging Face, runnable via SGLang-Omni, Diffusers, or ComfyUI.
Why it matters
Most open music models have been limited to short loops or 30-second clips. Music 3.0's five-minute coherence window and explicit structural controls address the main practical gap for creators who need full songs with consistent vocal identity, evolving arrangements, and professional mix quality. The Structured Caption system also lowers the barrier for non-musicians to specify arrangement-level intent without specialist vocabulary.
Our take
The hidden-state fusion approach is the key differentiator: by passing continuous LLM representations directly into Flow Matching rather than relying on discrete RVQ tokens for synthesis, MiniMax preserves acoustic nuance that token-only decoders typically lose over long sequences. Whether the 8B Global LLM's Qwen3 initialization transfers enough musical reasoning to rival proprietary systems like Suno v4 or Udio remains to be benchmarked, but the architecture is sound and the open release invites rapid community iteration.
