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 capable of generating complete five-minute songs at 32 kHz, 16-bit stereo WAV quality in a single pass The Hybrid-LM architecture splits work between an 8B Global LLM (long-range structure) and a 0.6B Local LLM (per-frame detail), feeding a 2.4B flow-matching synthesis module and a 123M Flow-VAE Unlike typical discrete audio models, synthesis runs on fused continuous hidden states, completely bypassing the discrete RVQ tokenize
Analysis
TL;DR
- MiniMax released MiniMax-Music3, an open-weights text-to-music model capable of generating complete five-minute songs at 32 kHz, 16-bit stereo WAV quality in a single pass
- The Hybrid-LM architecture splits work between an 8B Global LLM (long-range structure) and a 0.6B Local LLM (per-frame detail), feeding a 2.4B flow-matching synthesis module and a 123M Flow-VAE
- Unlike typical discrete audio models, synthesis runs on fused continuous hidden states, completely bypassing the discrete RVQ tokenizer decoder at inference time
- Deployment is practical: runs on two GPUs via SGLang-Omni, under 24 GB VRAM with diffusers, or as low as 8 GB with group offloading
- Commercial use is permitted under the Community License with visible attribution; organizations exceeding $20M annual revenue require separate written authorization
Why It Matters
MiniMax-Music3 represents a significant step toward production-ready AI music generation by combining open-weights accessibility with deployable infrastructure on day one, lowering the barrier for creators and studios to integrate high-fidelity music generation into their workflows. The hybrid architecture and continuous synthesis approach demonstrate a practical path toward long-form, structurally coherent music generation that balances quality with computational efficiency.
Technical Details
- Hybrid-LM Architecture: An 8B Global LLM (initialized from Qwen3-8B/Qwen3.5-8B) handles frame-by-frame prediction of the semantic RVQ codebook and maintains long-range musical structure, while a 0.6B Local LLM predicts the seven residual acoustic codebooks within each frame
- Training Tokenizer: Eight-layer residual vector quantization (RVQ) with a 16,384-entry semantic codebook and seven 1,024-entry acoustic codebooks; training optimizes the semantic layer first, then all eight jointly
- Continuous Synthesis Stack: Final hidden states from both LLMs are fused and conditioned on a 2.4B flow-matching module that maps into a latent space decoded by a 123M Flow-VAE (inherited from MiniMax Speech); the discrete tokenizer decoder is not loaded at inference
- Two-Input Control System: Lyrics with section tags ([Intro], [Verse], [Chorus], etc.) are provided separately from a Structured Caption containing Global Metadata, Vocal Details, and Arrangement; a music-caption-rewriter agent skill expands short descriptions into the required format
- Three Serving Paths: SGLang-Omni (reference server, two CUDA GPUs), diffusers modular pipeline (under 24 GB VRAM, ~22 GB with CPU offload, 8 GB with group offloading), and ComfyUI with native Text to Music template using FP16/INT8 weights
Industry Insight
- The open-weights release with day-one deployability signals a shift from research previews to production-grade AI music tools, enabling solo creators, indie studios, and mid-market teams to integrate music generation directly into pipelines without vendor lock-in
- The $20M revenue threshold in the Community License creates a clear segmentation: small-scale commercial use is accessible, while larger enterprises will need to negotiate terms, potentially driving adoption of alternative solutions or custom deployments at scale
- The two-input control design (lyrics + structured caption) with an agent-based rewriter establishes a pattern for controllable generative music that could become a standard interface for creative tools in gaming, advertising, and content creation workflows
Disclaimer: The above content is generated by AI and is for reference only.