MiniMax Music 3 Targets Full-Song Generation

The important shift is accessibility: high-quality, long-form music generation is moving from closed creator apps into developer-controlled workflows, though the hardware requirements remain substantial.

• –Hybrid global and local language models improve long-range song structure and acoustic detail
• –Supports section-aware lyrics, backing vocals, arrangement changes, instrumentation, and vocal direction
• –Open weights enable local deployment, customization, and integration into creative applications
• –Local inference currently requires two CUDA GPUs and supports non-streaming generation only
• –Prompt controls remain probabilistic, so tempo, key, lyrics, and arrangement may not precisely match instructions