Next upHack for Humanity: San Francisco (powered by Google Gemini)
News

MiniMax open-sources Music 3.0, a text-to-music model that writes five-minute songs

MiniMax released MiniMax Music 3.0, an open-weights text-to-music model that generates complete songs up to five minutes long from lyrics and a text prompt.

D
Aug 13, 2026 · 1 min read

MiniMax published MiniMax Music 3.0 on Aug 13, 2026, an open-weights text-to-music model that generates complete songs up to five minutes long from lyrics plus a text description.

The release extends MiniMax’s push to ship open-weights multimodal systems rather than keep them behind an API. Music 3.0 outputs 32kHz, 16-bit stereo audio and reads structured song-section tags such as [Verse], [Chorus], [Bridge] and [Outro], letting users lay out a track’s arrangement in the prompt.

Under the hood, the model splits the work across four stages, according to MiniMax’s technical write-up. An 8-billion-parameter “Global LLM” handles long-range song structure, a smaller 0.6-billion-parameter “Local LLM” fills in frame-level acoustic detail, a 2.4-billion-parameter flow-matching stage refines the audio, and a 123-million-parameter Flow-VAE decoder produces the final waveform. Here, LLM refers to a large language model adapted to generate audio tokens rather than text.

The weights are available on Hugging Face, GitHub, ModelScope and ComfyUI 0.33.0 or later, under the MiniMax-Music3 Community License. The model lands alongside MiniMax’s H3 omni-modal video model, first published Aug 3, a 33-billion-parameter system that generates short 2K video clips with native stereo audio. Together with an H3-Turbo variant, the three models were among the week’s trending models on Hugging Face.

Two caveats temper the “open” label. The H3 license bars local deployment in the US, EU, UK and South Korea, and MiniMax has published no independent evaluation of Music 3.0’s audio quality against rival music generators.

More news