MiniMax Releases Open-Weight Music Model MiniMax-Music3
MiniMax released MiniMax-Music3, an open-weights model generating five-minute songs from lyrics and structured captions, with commercial licensing terms.
The weights, inference code and three documented serving paths were published on the same day, making the model deployable immediately rather than a research preview. The MiniMax-Music3 Community License permits commercial use, but requires prominent display of “MiniMax-Music3” in the product UI. Organizations whose aggregate annual revenue from those products exceeds US$20 million must obtain prior written authorization from MiniMax. Hosting third-party generation also requires safeguards against infringing outputs.
The architecture pairs a Hybrid-LM—an 8B Global LLM and a 0.6B Local LLM—with a continuous synthesis stack built on flow matching and a Flow-VAE. The training tokenizer uses eight layers of residual vector quantization, with a semantic codebook of 16,384 entries and seven acoustic codebooks of 1,024 entries each. Training optimizes the semantic layer first, then all eight jointly. The Global LLM predicts the first RVQ codebook frame by frame and holds long-range structure; the Local LLM predicts the remaining codebooks within each frame. Rather than decoding from discrete RVQ tokens, the synthesis stage fuses final hidden states of both LLMs and conditions a 2.4B flow-matching module, whose latent output is decoded by a 123M Flow-VAE inherited from MiniMax Speech. At inference, the discrete tokenizer decoder is not loaded.
The model card and license state the Global LLM is initialized from Qwen3-8B, while the MiniMax Research post says Qwen3.5-8B, so the exact base checkpoint is unsettled.
Three deployment paths are documented. SGLang-Omni is the reference server, requiring two CUDA GPUs: GPU 0 runs Qwen3 and RVQ autoregressive generation, GPU 1 runs flow matching and decoding. A diffusers modular pipeline fits under 24 GB VRAM at full precision, about 22 GB with CPU offload, and down to 8 GB with group offloading. ComfyUI offers a native Text to Music template using repacked FP16/INT8 weights.
The report lists applications including background scoring for UGC video, adaptive game and level music, localized advertising and sonic branding, scratch and demo tracks for songwriters, mood-conditioned playlist generation, and offline batch generation. Potential users include solo creators, indie studios, and mid-market teams.