MiniMax open-sources Music 3, a Qwen3-based song generator

MiniMax has open-sourced MiniMax Music 3, a text-to-song model that generates complete tracks up to five minutes long from lyrics plus a music description. The output is 32 kHz, 16-bit stereo WAV, and lyrics can carry section tags like [Intro], [Verse], [Pre-Chorus], [Chorus], [Bridge], [Instrumental], [Solo] and [Outro] to shape structure.

The stack is not a single monolith. A hybrid language-model system pairs an 8B Global LLM that 'predicts the first RVQ codebook frame by frame and models the song's long-range semantic and structural progression' with a 0.6B Local LLM that fills in the remaining acoustic codebooks. The Global LLM is initialized from Qwen3-8B. Their hidden states are fused and passed through a 2.4B Flow Matching stage into a 123M Flow-VAE decoder. The tokenizer is eight layers of residual vector quantization, with a 16,384-entry semantic codebook and seven 1,024-entry acoustic codebooks.

Control is textual on both sides. Lyrics carry the words and the section markers; a separate music-description field is meant to cover 'musical style, emotional progression, vocal performance, instrumentation, arrangement, and production profile.' MiniMax concedes those descriptions 'provide generative control rather than strict symbolic guarantees,' so a request for a specific tempo, key or instrument may not land exactly.

Practical use has hard edges. Inference requires CUDA, only non-streaming generation is supported, text prompts cap at 5,000 tokens, and audio generation is limited to 9,000 acoustic frames. The repo ships example code for both a Diffusers pipeline and a SGLang-Omni server, and it publishes no benchmarks and no description of the training corpus. Two experts in our Who's Who directory have already shared the repo.

For indie musicians and downstream tool builders, though, an open-weights, full-song, section-aware model that runs on local hardware is the doorway that hosted music APIs have not opened.