MiniMax Drops Music 3, an Open-Weight AI That Writes 5-Minute Songs
MiniMax releases an open-weight text-to-music model that generates full 5-minute songs in 32 kHz stereo with native ComfyUI support.
- MiniMax released Music 3, an open-weight text-to-music model generating full 5-minute songs at 32 kHz stereo.
- Architecture pairs an 8B Qwen3-based global LLM with a 0.6B local LLM and Flow-VAE synthesis.
- Native ComfyUI support via the Comfy-Org repack, runs on ~22 GB VRAM with offloading.
- Commercial use allowed with attribution; separate agreement only required above $20M revenue.
- Quality compared to Suno V3.5, strong on vocals and lyric timing, weaker instrumental texture than V5.
- GGUF and MLX MXFP4 quantizations already available for low-VRAM and Apple Silicon setups.
MiniMax has released MiniMax Music 3, an open-weight model that generates structured songs from section-tagged lyrics and a written description of the sound. It produces up to five minutes of 32 kHz stereo audio, including vocals, instrumentation, and transitions between song sections. The ComfyUI package had passed 950,000 Hugging Face downloads when this article was prepared, although that counter includes file requests rather than unique installations.
Downloadable weights give developers control over deployment, quantization, fine-tuning, and data handling without tying production to a hosted API. Open-weight access still comes with a custom license, hardware costs, and responsibility for operating the generation pipeline.
The release at a glance
| Capability | Details |
|---|---|
| Inputs | Tagged lyrics and a prose music description |
| Maximum duration | Five minutes, with earlier stopping possible |
| Audio format | 32 kHz stereo |
| Core models | 8B Global LLM and 0.6B Local LLM |
| Primary runtimes | ComfyUI, Python with Diffusers, and community MLX or GGUF builds |
| Approximate VRAM | About 22 GB with CPU offloading, or 8 GB with additional group offloading |
A song stack in four stages
Music 3 uses a hierarchical autoregressive architecture, meaning each new segment depends on earlier segments while separate models work at broad and fine time scales. This division lets the system track a song’s structure across minutes and restore acoustic detail within individual frames.
- Global planning: An 8B Global LLM initialized from Qwen3-8B tracks lyrics, themes, rhythm, vocal identity, and arrangement across the song.
- Local detail: A 0.6B Local LLM models short-range acoustic information within each frame.
- Latent synthesis: A flow-matching transformer converts the language-model states into Flow-VAE latents one chunk at a time.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
