MiniMax-Music3 open-weights open a production-ready path for AI-generated full songs

MiniMax released MiniMax-Music3, an open-weights text-to-music model that generates a complete song up to five minutes from lyrics and a structured music description. Unlike many AI music platforms that remain closed, MiniMax shipped weights, inference code, and three documented serving paths on day one, making this deployable now rather than a research preview.

A hybrid architecture for long-range coherence

MiniMax-Music3 uses a Hybrid-LM design: an 8B Global LLM predicts the first RVQ codebook frame by frame and holds long-range structure, while a 0.6B Local LLM predicts the remaining codebooks within each frame. The model card states the Global LLM is initialized from Qwen3-8B, though a MiniMax Research post says Qwen3.5-8B; 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.

Inputs are split into two parts: lyrics with section tags ([Intro], [Verse], [Chorus], etc.) and a separate Structured Caption carrying Global Metadata, Vocal Details, and Arrangement. MiniMax also ships a music-caption-rewriter agent that expands a short description into that three-part format offline.

Practical deployment: hardware and serving paths

Three documented serving paths shipped on day one. SGLang-Omni is the reference server, requiring two CUDA GPUs: GPU 0 runs the Global LLM and RVQ autoregressive generation, GPU 1 runs 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.

For builders, this means you can integrate MiniMax-Music3 into a production pipeline today, either via SGLang-Omni for custom serving or through ComfyUI for visual workflow prototyping. The open-weights release includes the model on Hugging Face and inference code on GitHub.

Caveats and open questions

The hardware requirements are non-trivial: two GPUs with specific assignments, and the diffusers path still needs at least 8 GB VRAM with aggressive offloading. The exact base checkpoint for the Global LLM is unclear, which may matter for reproducibility or fine-tuning. No independent audio quality benchmarks or comparisons to closed models like Suno or Udio are available in the provided sources. Licensing terms are not detailed in the release description; check the Hugging Face model page for the specific license.

Despite these caveats, MiniMax-Music3 represents a significant step: a production-ready, open-weights music generation model that can generate full songs with structural coherence, backed by a well-documented architecture and multiple deployment paths.