MiniMax released MiniMax-Music3, an open-weights text-to-music model capable of generating complete five-minute songs from lyrics and a structured caption. The system outputs 32 kHz, 16-bit stereo WAV files in a single generation. Its architecture combines a Hybrid-LM, consisting of an 8B Global LLM and a 0.6B Local LLM, with a continuous synthesis stack based on flow matching and a Flow-VAE. Weights, inference code, and three documented serving paths were released on the same day.
Deployment readiness
The company published usable weights and inference code immediately, making this a deployable tool rather than a research preview. The MiniMax-Music3 Community License allows commercial use but requires prominent display of the model name in product UI. Any organisation with aggregate yearly revenue exceeding US$ 20 million must obtain separate prior written authorization from MiniMax. Hosts of third-party generation must also implement safeguards against infringing outputs.
Target industries include game development, advertising, short-form video, e-learning, podcasting, and music-tech SaaS. Specific applications cover background scoring for user-generated content, adaptive game music, localized ad beds, scratch tracks for songwriters, and offline batch generation.
System design
The training tokenizer uses eight layers of residual vector quantization (RVQ). The first semantic codebook contains 16,384 entries for core musical structure, while the remaining seven acoustic codebooks hold 1,024 entries each for residual detail. Training optimises the semantic layer first, then all eight layers jointly.
The Hybrid-LM splits the modelling task. An 8B Global LLM predicts the first RVQ codebook frame by frame to manage long-range structure, while a 0.6B Local LLM predicts the remaining codebooks within each frame. Documentation states the Global LLM is initialised from Qwen3-8B, though a research post mentions Qwen3.5-8B; treat the exact base checkpoint as unsettled.
The synthesis stage fuses the final hidden states of both LLMs to condition a 2.4B flow-matching module. This 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.
Input control
Lyrics carry words and section tags on separate lines, including markers for Intro, Verse, Pre-Chorus, Chorus, Post-Chorus, Bridge, Instrumental, Solo, and Outro. A separate Structured Caption holds Global Metadata, Vocal Details, and Arrangement. MiniMax also ships a music-caption-rewriter agent skill that expands a short description into this three-part format offline.
Running the model
SGLang-Omni serves as the reference server. The GitHub page specifies two CUDA GPUs: GPU 0 runs Qwen and RVQ autoregressive generation, while GPU 1 handles flow matching and DAV decoding. The diffusers modular pipeline fits under 24 GB VRAM at full precision, drops to about 22 GB with automatic CPU offload, and reaches 8 GB with leaf-level group offloading. ComfyUI offers a native Text to Music template using repacked FP16/INT8 weights from Comfy-Org.
What it means
Creators can now generate full-length tracks in one pass without needing to stitch together short segments. The ability to run on consumer-grade hardware via group offloading lowers the barrier for indie developers and solo producers who previously relied on expensive cloud APIs.
