MiniMax Releases MiniMax-Music3: An Open-Weights Music Model Generating Complete Five-Minute Songs From Lyrics and a Structured Caption

MiniMax released MiniMax-Music3 , an open-weights text-to-music model. The model takes two separate inputs: lyrics carrying section tags, and a detailed music description. It returns a complete song of up to five minutes in a single generation, as 32 kHz, 16-bit stereo WAV. The architecture pairs a Hybrid-LM, an 8B Global LLM with a 0.6B Local LLM, with a continuous synthesis stack built on flow matching and a Flow-VAE. Weights, inference code and three documented serving paths shipped the same day.

Yes, MiniMax published usable weights, inference code and three documented serving paths on day one, so this is deployable now rather than a research preview.

MiniMax-Music3 combines a hierarchical autoregressive stack with a continuous synthesis path.

The training tokenizer uses eight layers of residual vector quantization (RVQ). The first, semantic codebook has 16,384 entries and carries core musical semantics and structure. The remaining seven acoustic codebooks have 1,024 entries each and encode residual detail. Training optimizes the semantic layer first, then all eight jointly.

The Hybrid-LM splits the modeling problem. An 8B Global LLM predicts the first RVQ codebook frame by frame and holds long-range structure; a 0.6B Local LLM predicts the remaining codebooks within each frame. The model card and license state the Global LLM is initialized from Qwen3-8B ; the MiniMax Research post says Qwen3.5-8B, so treat the exact base checkpoint as unsettled.

The synthesis stage is the more interesting design choice. Rather than decoding from discrete RVQ tokens, MiniMax fuses the final hidden states of both LLMs and conditions a 2.4B flow-matching module on them, which maps into a latent space decoded by a 123M Flow-VAE inherited from MiniMax Speech. At inference the discrete tokenizer decoder is not loaded at all.

Lyrics carry the words and section tags on their own lines: [Intro] , [Verse] , [Pre-Chorus] , [Chorus] , [Post-Chorus] , [Bridge] , [Instrumental] , [Solo] , [Outro] . A separate Structured Caption carries Global Metadata, Vocal Details and Arrangement. MiniMax also ships a music-caption-rewriter agent skill that expands a short description into that three-part format offline.

Three documented paths. SGLang-Omni is the reference server; the GitHub page specifies two CUDA GPUs, with GPU 0 running Qwen3 and RVQ autoregressive generation and GPU 1 running flow matching and DAV decoding. The diffusers modular pipeline fits under 24 GB VRAM at full precision, about 22 GB with automatic CPU offload, and down to 8 GB with leaf-level group offloading. ComfyUI has a native Text to Music template using repacked FP16/INT8 weights from Comfy-Org .