MiniMax Drops MiniMax-Music3, an Open-Weight AI That Writes Full Songs

MiniMax has released MiniMax-Music3, an open-weights text-to-music model that turns a lyric sheet and a written sound description into a complete song. It takes two separate inputs, lyrics carrying section tags and a detailed music description, and returns up to five minutes of 32 kHz, 16-bit stereo WAV in a single generation. Weights, inference code, and multiple serving paths shipped on day one, which is unusual for a model of this scope.

Open weights have been catching up quickly in this space. MiniMax-Music3 arrives just weeks after ACE-Step 1.5 claimed to beat Suno v5 on standard benchmarks, and months after YuE showed that open-weight models could match commercial systems in human listening tests. What sets this release apart is that the whole stack, including a Qwen3-based language backbone and a flow-matching audio decoder, is downloadable.

How the model is wired together

The architecture is hierarchical and autoregressive, split into a model that thinks about the whole song and a smaller one that fills in acoustic detail. The Hybrid-LM framework combines an 8B Global LLM for long-range structural modelling with a 0.6B Local LLM for frame-level acoustic details. The synthesis pipeline then fuses the continuous hidden states of both LLMs to condition a 2.4B flow-matching module and a 123M Flow-VAE decoder, bypassing the discrete synthesiser decoder during inference.

The Global LLM is not built from scratch. Initialized from Qwen3-8B, it models the song's long-range semantic and structural progression, while the 0.6B Local LLM restores fine-grained acoustic detail within each frame. Instead of decoding audio from discrete tokens alone, the synthesis module fuses the hidden states of both LLMs. That last detail is the interesting bit. Most token-based music systems throw away the LLM's rich internal representations and reconstruct audio from quantized codebook indices, which caps the fidelity of what the decoder can recover. Passing continuous hidden states directly into flow matching gives the decoder more to work with.